Machine Vision 4 Users

Accessing the inner circle

2012-02-01T22:45:00.000-05:00

Newbies to the world of machine vision – and I count myself in their number – often find there is a mysterious inner circle of experts distinguished from we lesser mortals by their ability to use cryptic words.

Chief among these words are “jitter” and “latency”. You’ll see them in framegrabber specs and hear them bandied around by those who were doing machine vision when Bill Gates was a boy, but if you don’t know what they mean, it’s really difficult to find out.

But thanks to Vision Systems Design, that glass ceiling has been shattered. “Exposing jitter and latency myths in Camera Link and GigE Vision systems” (January 1^st 2011) summarizes a Dalsa paper on the issue. One interesting point to note is that the author of the Dalsa paper concedes there are variations of the definitions, which is guaranteed to make like complicated.

If you want to understand jitter and latency, and I suggest you should, follow the link and become a guru of machine vision. Or at least be able to talk like one.

An alternative to laser triangulation

2012-01-30T22:24:00.003-05:00

It seems like everyone wants to do 3D imaging, and for most the method of choice is laser triangulation. It’s relatively simple to implement, the processing doesn’t need a supercomputer, and it needn’t cost a fortune, but it does require some relative motion between laser line and object under inspection.

Enter Time-of-Flight imaging. The basic concept has been around a long time; I think Canesta (subsequently acquired by Microsoft,) was one of the first companies to demonstrate a ToF sensor and Fotonic still seems to be in the market, but it hasn’t got the airtime given to laser-based methods.

Odos Imaging ltd of Scotland is out to put that right. They’ve launched a “unique 2D+3D™” camera that, as their description implies, captures both a conventional grayscale image AND distance information for each pixel of the image.

If you visit their website you’ll see some images that show what the camera can do. The reason this interests me is that it eliminates the need to scan a surface: there is no need for relative motion. I see this being especially useful in autonomous navigation applications, which is an aspect of machine vision with growth potential.

No word on pricing, but I’m looking forward to reading the first application story in Vision Systems Design.

Three types of computer vision companies?

2012-01-29T22:31:00.000-05:00

Denver-based Boulder Imaging argue that you can buy your machine vision technology from the component manufacturers, from a vision system integrator, or from a vision solution provider. I think their analysis is oversimplified, but why don’t you read their piece and make up your own mind? “Did You Know... There Are 3 Types of Computer Vision Companies?”

So why do I think this is a simplification?

What Boulder are overlooking is the experience of the integrator plus their use of off-the-shelf components. The best integrators have been around a long time and have great depth of experience. Plus, utilizing standard hardware and software means repair, support and upgrading is relatively straightforward.

Of course, not every integrator does a top-notch job on every project, but when things go “pear-shaped” as one of my colleagues likes to say, it’s my belief that the customer is at fault.

Yes, yes, the customer is always right, but often they don’t do a good job of managing the integrator. In fact I’d go further and suggest that many buyers of vision integration services don’t even realize they should be managing their integrator.

Wouldn’t that make a great theme for a series of blog posts?

Cables, the Rodney Dangerfield of machine vision?

2012-01-26T22:26:00.000-05:00

Cables – those plastic-coated wires that enable your vision system to do something useful – can’t get no respect. They’re bent, twisted and folded to fit in the tightest of spaces, and the connectors …. They get shoved in and pulled out like a … well I had a crude analogy in mind but I shan’t print it here. You know what I mean.

If you want your vision system to be robust and industrial, (and who doesn’t?) then you need to look after your cables, AND use industrial connectors. Thumbscrews are good, but what I really like are the M12-type connectors I’ve seen on the latest cameras from Matrox and NI.

And according to “Connectors and Cables 101” (another Vision & Sensors article, January 4^th, 2012,) I am not alone in this. Click the link to learn what Intercon 1 has to say on the subject.

New smart cameras from NI

2012-01-24T22:28:00.002-05:00

Here’s a puzzle. Why does the NI website show two different products as a NI 1772 Smart Camera?

Seriously. Here’s the old, strangely-styled 1722 priced at $1,999, and here’s the chunky new 1722, priced at $3,598.

That’s quite a bump in price, so what’s going on? Well the new cams, billed as “High-Performance” and running from VGA to 5Mp, use a 1.6GHz Atom processor, while the old – and presumably outgoing – model harnesses a 400 MHz PowerPC processor. That’s quite a bump in horsepower, but does it really warrant a $1,500 price increase?

Or, have NI just noticed that competing products, from the likes of say, Matrox, are priced from around $3,500 for an Atom-powered, VGA res smart camera, so they figured they should position their camera at the same level?

Or, could there be more to this than meets the eye? If you’re interested in my hypothesis, check back in a day or two.

What’s in a name?

2012-01-23T22:04:00.003-05:00

The Bard told us “that which we call a rose would by any other name smell as sweet,” so do titles matter?

Jeff Burnstein of the Automated Imaging Association (now just the AIA,) seems to think so, because they are dropping the word “machine” from their website, going from “machinevisiononline.org” to just “visiononline.org”.

Why the change? Well, as Burnstein says, “We see the technology extending to every corner of the globe and into every industry, well beyond the confines of the factory.” I agree that many of the advances we’re seeing today are in fields such as agricultural automation and vehicles guidance – far from the realm of industrial inspection – so this name change is really nothing more than a recognition of reality. And since the AIA is a trade organization representing those who make and sell cameras, lights, optics and so on, they don’t much care if the products go into factories or greenhouses, so long as they go.

But is “vision” the best term to use? Should we perhaps speak of “automated vision” to differentiate it from human vision? I agree with dropping “machine” as that implies industrial applications, but simply “vision” is, in my humble opinion, too broad.

Can a worse image produce better results?

2012-01-22T22:07:00.000-05:00

Scanning my October ’11 edition of Vision Systems Design, I came across an interesting article by Stuart Singer of Schneider Optics. Nothing unusual about that – everything Stuart writes is interesting – but his articles do tend to be a little intense, so I set it aside for later reading.

Now most of what he says in “Filters Factor into Optical Imaging” goes over my head, although I think I get the main point: a lens filter will affect the MTF of your optical system. But it was a throwaway line about Nyquist theory and limiting resolution at the sensor that really caught my eye.

What Stuart said was that the smallest object that can be resolved by a pixel of 7 µm is 14 µm, according to the Nyquist theorem. This sent me off on a quest to learn more about Harry Nyquist and his work. I stumbled across a great definition here, but the most interesting find was buried in some notes from the Sensory Processes class at the University of Berkeley. On the last page of this six pager about Aliasing are four photos illustrating how blurring an image can have the effect of adding more information.

I’m oversimplifying, I know, but it set me wondering if a low pass filter would improve the performance of a precision gauging application I’m working on. So I grabbed a set of images under slightly varying lighting (just like in the real world,) and made some measurements. I got a variation between images of around 4 pixels.

Then I applied a low pass filter to the images and repeated the measurement. This time the variation was 0.4 pixels. In others, and I’m speaking crudely, blurring the image, like in the Berkeley paper, produces a more stable result.

Math is not my strong suit, so I’m not going to attempt an elegant explanation of why a worse image produces a better result, but I would welcome comments on and explanations of my findings.

Open source machine vision software?

2012-01-19T22:34:00.000-05:00

We know there are are two certainties in life: death and taxes. Well there’s a machine vision equivalent: software is complex and expensive. For better or worse, that whole open-source, shareware thing doesn’t seem to have penetrated our specialist world. But there are signs of change upon the horizon.

SimpleCV “is a Python interface” from Ingenuitas that purports to be not only simple, but free. I swapped emails recently with one of the brains behind Ingenuitas to find out more. Here’s what Nate Oostendorp had to say:

“…I was curious about machine vision and what the state of the art was, and I started looking around at Open Source software for doing machine vision programming. OpenCV seemed to be the defacto choice for doing vision work, but the learning curve for it was very high -- particularly for myself, who hasn't had any formal training in machine vision -- just experience from the practitioner side. Any time I saw methods described with equations or algorithms referred to by their discoverer -- rather than what they do -- i found it really pretty difficult to pick up and use, starting from scratch.

“So my co-founders and I formed Ingenuitas [and] started building SimpleCV, with the goal of providing a framework for vision applications -- in the same milieu as Rails or all the frameworks that are out there for the web. A set of tools to get started quickly, write as little code as possible, and make the simple stuff really really easy.

“We're working now on version 1.3, and planning for what is going to go into 2.0 -- we're also building an Open Source web-based GUI for machine vision called SimpleSeer. Our goal is to make setting up a vision appliance as easy as possible and letting people use the type of commodity equipment (DSLR cameras, webcams, Kinects) that they're already comfortable with. The sensors are out there, and computers have gotten (and will continue to get) faster and cheaper -- we're hoping we're the glue that can bring these things together and make cameras useful for as many things as possible.”

So that’s what SimpleCV is all about. If it lives up to its promise it could be something of a paradigm shift, but first I think we need to wait for the next release. Hopefully, I’ll have news on that soon, so keep checking back.

What can multi-spectral imaging do for you?

2012-01-18T22:16:00.000-05:00

I’ve long been fascinated by the 2-CCD cameras from JAI. I know they do a 3-CDD version too, and there’s an interesting application case study on the Vision & Sensors site (Jan 5^th, 2012,) but that’s aimed more at color applications. What I like about the 2-CCD model is that it combines visible and IR imaging through the same lens, for which I can see many possible applications. (Okay, full disclosure: I haven’t bought any of these because they’re somewhat expensive, but the concept is great.)

For those of you who struggle to see where you would use such a camera, JAI have placed a short movie on YouTube, but to save you time I’ve embedded it below.

What has always intrigued me though is that JAI seem to miss what I see as the biggest benefit of this camera concept: it allows the use of multiple lighting strategies at the same location on the production line.

For example, suppose I need both a darkfield image and a front-lit image of a part moving down a production line. I’m forced to have two separate stations with two cameras, and that takes up valuable floorspace. With an AD-080CL camera I need only one station, saving me precious space.

Now I shall sit back and wait for my royalties from JAI.

Tell me and I will forget, show me and I will remember

2012-01-17T22:47:00.000-05:00

Well I think that’s how the quote goes. The point is that most of us learn better when we’re able to actually do things for ourselves rather than just reading about them. This is why kids hands-on museums are so full of adults playing with the exhibits and ignoring their precious little ones.

But rather than moan about other people’s parenting skills, today I wish to share a link to imageprocessingbasics.com. This has some fantastic tutorials on various aspects of image processing, such as histogram equalization and convolution. To make life simple though, each tutorial comes with a Java applet that allows the viewer to see the effect of altering various imaging parameters. It certainly worked for me, so give it a go.

By the way, thanks are due to the guys at Vision Systems Design for bringing this site to my attention. Thanks guys.

I’m so pedestrian

2012-01-16T22:26:00.000-05:00

Most of my machine vision applications involve pretty low data rates. It’s rare that I need to move more than a 2Mp image at anything faster than 15fps, so the flurry – no, make that a torrent – of new camera standards has kind of washed over me.

The latest to leave me bemused is GigE Vision 2.0. I’m still just finding my feet with GigE (and I do like it,) but I didn’t realize version 2.0 was even in development. The big deal is higher data rates, along with better support for Link Aggregation, (as used on the Prosilica GX range from AVT,) but there are some other enhancements that will be of interest to the geekier among you. If you’re looking for more info, I suggest you start with “Gige Cameras: GigE Vision 2.0 Has Arrived” on the Vision & Sensors website, January 10^th, 2012.

This won’t tell you all you need to know, but I think it’s a good start point.

Shouldn’t autofocus be standard?

2012-01-12T22:38:00.000-05:00

I make it a requirement that I have remote access to all my installed vision systems. This saves me much time and money in travelling to factories to make a simple change in the way a system works. However, there remain two classes of problem that still require site visits: lighting and optics.

If you’re wondering why, I take it you’ve never tried to advise a technician at the end of a phone on how to focus an image, while watching the result live over the internet. “Turn the focus ring left …. No, that’s the aperture. The other ring … no, the other way …. No, turn the ring the other way. Too far…”

Autofocus lenses exist and zoom technology is commonplace in traffic control, (AVT have a solution,) but we haven’t really adopted them in machine vision. I think it’s time that changed. Of course, this will force a shift in camera design since the commands to the lens will pass through the camera, but I don’t see any real obstacles to doing that.

Not only would this help maintenance, it would also mean the lens could become a change part. If a system needed one field-of-view for a large part and a much smaller FoV for a second, then why not just change lenses? At the moment I would never propose doing this because the focus will never go back the same way twice, but with autofocus that problem goes away.

That just leaves lighting problems that need a site visit. Brightness I can control through software, but the angle? Has anyone considered applying optics technology to lighting?

Feedback on evaluating LED lights

2012-01-10T22:11:00.000-05:00

Back in November I posted a few words on wisdom on how I go about assessing the quality and performance of LED machine vision lights from the many vendors out in the marketplace. The point I was trying to make was that most of us lack any sophisticated test equipment, so instead we have to make do with a few more subject methods.

Well the guys at Microscan took the time to write a detailed comment to my post, but as I’m not sure how many readers will actually go back to look at that article I thought I’d take the liberty of reproducing their comments here.

Great blog post! I asked our NERLITE veteran engineers for their perspective; here is their reply:

While it is a fact that the cooler a particular LED is the longer it will last, this is much too broad of a generalization. Different LEDs have a very wide range of maximum operating temperatures. If the criteria is “hot to the touch”, some LEDs may already be past their limits while others might not be even close. This used to be a good "rule of thumb" in the early days of LED illumination, but devices have significantly evolved and heat does not always mean a bad design. For example, Microscan Nerlite’s new Smart Series DOALs run hotter than our previous design and are warmer to the touch, but the LED's are robust enough that we were able to increase intensity, improve uniformity, eliminate cooling fins and operate the unit at a 10 degree C greater maximum ambient temperature.

We agree [regarding the need for good mechanical rigidity and mounting points]. For example with the Smart Series product line, we were able to take into consideration feedback from application engineers and customers to incorporate more common mounting points while maintaining backward mounting compatibility to prior product. Each design was subjected to shock and vibration testing to ensure mechanical stability in the field and we also took into consideration IP ratings that are common to each product family. In terms of cleaning, it is important to note all units with beam splitters, regardless of manufacturer, have special cleaning requirements. This is due to the delicate nature of the coatings used on the beam splitters.

We agree [regarding the importance of LED placement uniformity]. If the LEDs are not positioned uniformly, the output will not be uniform either. With non-diffuse, focused LED arrays, the trueness of LED placement makes a big difference and dictates uniformity of light coverage. For example when Nerlite products require specific LED alignment, we either incorporate the alignment feature into the design, or specifically design alignment tools to be used at the time of assembly.

[On how to assess uniformity] For non-diffuse, short working distance area arrays and ring lights, this is basically true. [Small LEDs are to be preferred over big ones.] It does not take into account the output angle of the LEDs’ optics. Even packed tightly together, an LED with a very narrow output angle can still be spotty. LEDs with a wide output pattern can be more uniform. This is true when considering the uniformity of non-diffuse, focused LED arrays, but when it comes down to having a truly diffuse light source, seeing "any" type of LED is not a good sign. Nerlite has applied many techniques, often patent protected, to achieve superior uniformity regardless of LED size or quantity. One final note is that for area arrays designed for long working distances (Hi-Brites), the overlap that occurs over the long distance makes LED density less of an issue.

[Assessing value-for-money – I suggested calculating the price per watt.] This may be true when comparing "apples to apples", but Machine Vision Illumination has entered a new era where more technical features are designed into the product itself. This simplistic approach also does not take into account the quality and integrity of construction, IP rating, optical design, efficiency of the electronics and optics, ease of use (mounting options, accessories, etc.), warranty, and so on. Even if all other factors are equal, some LEDs are more efficient than others. For example, the new smart rings use about the same amount of power as their predecessors, yet they contain 25-33% more LEDs and have 3-5X the light output depending on the model. Since the price and power consumption is about the same, by this criteria, one would judge the old and new units to be equal. This is, of course, totally false. Whether it's just simple circuit protection or the comprehensive control features built into Smart Series Illuminators, there are vast differences and sometimes you truly do get what you pay for.

This is great feedback, although I’m not sure how many of these are ‘actionable points’ that can be applied by the machine vision end-user. I think my main point is still valid: it’s really difficult to tell what makes one light worth double the price of an identically-sized product.

Keep those comments coming!

Off-the-shelf backlighting product

2012-01-09T22:18:00.000-05:00

Backlighting is one of the oldest machine vision lighting techniques, though it’s not without its challenges. Cylindrical metallic surfaces – think Coke can – can be extremely difficult to gauge in a backlighting setup and really need collimated light plus telecentric imaging.

Well it seems Keyence have identified this as a niche in which they didn’t have a presence, until now. Writing on SensorTips, (December 26^th, 2011,) Randy Frank describes the TM-3000 2D Dimensional Measurement System, which seems to include those aforementioned essential components. Randy also provides a link to the Keyence product page, though I suspect you will have to register for access.

Now I should mention that I’m pretty sure I saw Spencer Luster of Light Works demonstrate a very similar system a few years ago. If I’m correct, the big difference was that Spencer had a larger field of view, which, depending on what you might need to gauge, could be useful. So by all means, invite the Keyence salesperson in to give you a demo, (they’re very good at that,) but in the interests of doing some due diligence, ask Spencer what he could do. He might just have a few tricks up his sleeve that Keyence haven’t considered.

Kinect machine vision application

2012-01-08T22:24:00.000-05:00

Regulars to these pages know that I’m pretty excited about the potential of Microsoft’s Kinect, in fact I’m working on an industrial application right now, but I haven’t yet seen any real applications.

That has all changed with the announcement of the “Collision-Proof” power wheelchair from GeckoSystems. This uses the Kinect to detect hazards and stop with wheelchair user/driver from hitting them. There’s lots of info and a couple of movies on the page I’ve linked to, so just click on that. I think you’ll agree that this seems like a great application for this inexpensive 3D imaging system.

Future vision

2011-12-28T10:48:00.000-05:00

The end of one year and beginning ofthe next is traditionally a time to take stock. Magazines, TV shows and yes,even blogs like this, are full of such articles. But as I’m too pushed for timeto compose my own, (yes, that’s code for “lazy”,) I’m going to link to andcritique a great article of the looking forwards genre, and that’s “MachineVision: The Future of Machine Vision” written by Ben Dawson and published on the Vision &Sensors site, November 28^th, 2011.

Ben makes some great points, most ofwhich I fully endorse. I especially agree with his plea for kinder and gentleralgorithms, because I believe it’s lack of ease-of-use that’s holding back theindustry. But to expand on Ben’s points, what I’d really like are softwaretools that seamlessly account for or tolerate changes in lighting. I want thesoftware to understand the difference between an edge and a shadow, and I wantit to deal with lighting that changes when someone opens an enclosure door.

Readers might point to PatMax fromCognex as such a tool, but I’m not convinced. Yes it does some of what I want,but I’m looking for it to be invisible to the user. I want the vision system tojust understand the difference between a part and its shadow. Is that too muchto ask?

One area missing from Ben’s paper isoptics. I would love to see smarter optics. I want a machine vision system thatcan automatically focus on what I put in front of it, and can stay focused evenif the working distance changes. My feeling is that liquid lenses will give usthis ability within a year or two. It can’t happen soon enough!

Last, Ben has a great product idea.He discusses a room painting app – just snap a picture of your room or houseand then recolor it in a paint of your choosing. I’ve actually tried thismyself using Photoshop so I think it’s a great idea. What it needs though is adatabase of the reflective properties of every surface in the image. It also needsto identify what each surface is, and apply those properties. Then it has todetermine the light source and model how that light will scatter from thesurfaces and how much will reach the camera.

It’s all do-able, it just needsprogramming. Cut me in for 10% please Ben!

Yogi Berra once said, “It’s tough tomake predictions, especially about the future,” and I’m sure Ben’s list, withadditions by me, will prove that true. Maybe, just maybe, there’s somegame-changing idea lurking around the corner that will shake things up in wayswe can’t even imagine.

Holiday Cheer

2011-12-22T22:24:00.000-05:00

I’ve been looking for something amusing with which to round out the year, and just in time, Edmund Optics have come up with a winner. I feel they won’t object to me lifting the video below from their YouTube Channel, providing of course that you watch all their other videos. As there’s a lot of great content, that shouldn’t be too arduous a task.

See you in January!

The Basler approach to product naming

2011-12-21T23:05:00.000-05:00

Britain’s Jeremy Clarkson recently described the MacLaren MP4-12C supercar as sharing its name with a photocopier, and his audience got the joke immediately. Sexy products like supercars should have sexy names: which would you prefer: Aventador or MP4-12C?

I think this also applies to more mundane products like machine vision cameras. “MQ022MG-CM” is an engineering designation that may do a great job of describing the product, but it’s unlikely to lodge in anyone’s cerebellum.

Basler seems to get this, which is why they name their cameras, using words such as Ace, Pilot, and now “racer”. AVT get it too, basing their naming strategy on fish – Pike, Marlin, Guppy and so on. These are all product names that will stick in the mind: when I’m asked to suggest a camera I may not recall the MQ0-12C or whatever it is but I will remember “Stingray” or “Racer”.

The “racer” is an interesting new product. Described as “cost-effective” (which I imagine means “cheaper”,) this is the start on a new linescan family. It’s built around a CMOS sensor and is available with either Power-over-CameraLink or GigE interfaces. The latter of course saves a few hundred bucks on a framegrabber, though whether that’s negated by a price premium over the PoCL version I couldn’t say.

What I will say though is that I’m likely to remember the Racer next time I’m asked to suggest a camera while I doubt the DLC-2048-48-CL (yes, I made that up,) will spring as readily to mind.

Aspect ratios and wine bottles

2011-12-20T22:33:00.000-05:00

Despite all the growth, machine vision is still a niche industry. We’re not big enough to have the semiconductor guys make sensors just for us (leaving the linescan sub-niche out of the discussion,) so instead we have to make do with what’s churned out for consumer applications. Most of the time this isn’t a problem, but there are times when I find I’m throwing away a whole bunch of pixels.

Example? Well imagine viewing a wine bottle from the side. This has an aspect ratio something like 16:4 so to get the whole thing on a typical 16:9 aspect ratio sensor I have to include a whole lot of empty space in my image. Wouldn’t it be useful to have a long thin sensor?

Well, while I was reading up on Ximea’s USB3 cameras I spotted such a thing. Their sexily-named MQ022MG-CM has a CMV2000 sensor of 2048 x 1088 pixels. Now this sensor may be used by other camera vendors too, I haven’t bothered looking. It just struck me, while downing a glass of merlot, that long, thin sensors would be a useful thing to have.

Size wars

2011-12-19T22:35:00.002-05:00

If you took the time to follow my links to USB3 camera manufacturers, you might have noticed something interesting. Point Grey have been very eager to boast of their Flea3 being “the world’s smallest and fastest USB 3.0 cameras…”). And at 29 x 29 x 30mm, it is pretty darned small. But along comes Ximea with their MQ series of “Ultra-compact USB 3.0 camera[s]”. Dimensions on this are 26.4 x 26.4 x 19.1mm, so it’s smaller.

Now I struggle to imagine an industrial application where the Flea would be too big, but it’s amusing all the same. For some folks, size is very important.

And that’s all I have to say about that.

Is your mailto: broken?

2011-12-18T22:24:00.000-05:00

I do most of my solutions shopping on line. I Google the problem I need to solve, browse a few websites, and if something looks interesting I click on the contact provided. It’s usually either sales@avisioncompany.com or info@myvisioncompany.com, and that lets me shoot off an email letting the company know I’m interested in their products.

And then what?

Well probably 4 times out of 5, nothing. Zilch. Nada. Silence.

This drives me mad. What is the point of providing a contact email address on your website if no one ever reads the email?

Time was, I’d wait a day or so and then think, “No one ever got back to me from Avisioncompany so let me find their website again and see if there’s a number I can call.” These days though, I’m too lazy, so if you don’t read and respond to my email, you won’t hear from me again.

I’m not going to do a name-and-shame because so many companies are guilty of this slapdash practice. But do yourselves a favor; send an email to the contact on your website and check that it goes through.

Believe me, a lot of companies are losing a lot of business because they don’t check the mailto: on their site.

"Home Made" Hypercentric Lenses

2011-12-15T22:25:00.000-05:00

Optics articles from occasional contributor Spencer Luster of LightWorks don’t come along very often, but they’re worth the wait. Here’s a great article on DIY pericentric imaging.

One of Light Works' main product lines is hypercentric (sometimes called pericentric) lenses that provide converging views. We offer high quality models for difficult and critical inspection tasks. Many times, however, the basic converging view geometry is of benefit, even with lesser image quality. With that in mind I'd like to quickly describe how you can set up such geometry with a conventional camera lens and a Fresnel lens.

Figure 1 shows a spool with defect examples such as a dent in the top flange, and a tear in the bottom. Wouldn't it be great if you could inspect both flanges simultaneously using one camera? Now you can! Order before midnight tonight! I mean, uh, just keep reading.

Figure 1

Figure 2 shows an arrangement of a combination of a camera lens and a Fresnel lens that makes up a hypercentric system.

Figure 2

Some things to note about this system are:

The Fresnel Lens has a focal length FL.
The Fresnel Lens is placed two FL distances away from the Iris of the Camera Lens. (You can estimate by eye the location of the Iris.)
When the Iris is small – say F/8 or more – the Camera Lens will only accept light that appears to come from a small Virtual Source that is two FL distances below the Fresnel Lens.

The result of this system is a converging view of the object. (See the pink dashed lines.) By itself many fun and useful inspections can be achieved. Nevertheless, our goal was to inspect the top and bottom flanges. As shown, the bottom flange can be inspected, but the converging view of the top means it's "back lit" by the spool itself – it can't be silhouetted to spot the dent defect.

One easy solution is to cut a hole in the center of the thin, plastic Fresnel lens. Figure 3 shows this.

Figure 3

Now place a diffuse light source below the spool. The result? Pure optical magic! The bottom flange inspection hasn't changed, and the top flange is now properly back lit.

Okay, that's the Readers' Digest version. In practice there are a few things to keep in mind:

Fresnel lenses are for imaging use in near monochromatic light only. Stick with single color LED illumination. Even then image quality will only be fair.
Most Fresnel lenses are designed to have their grooved side facing a so-called "infinite conjugate." That is, the grooved side should be viewing something very far away, or receiving nearly collimated light. The smooth side should face a near object or the finite conjugate. Using a Fresnel as described above violates this principal, but you can still build a useful system. An improvement would be to use a pair of Fresnel lenses with their grooved sides facing each other.
The distance from the Fresnel to the Camera Lens Iris doesn't have to be 2* FL, but it's a good starting point. Experiment with other distances.
When illuminating the object, you may see light scattered from the Fresnel grooves. Try to prevent light traveling directly from the light source to the Fresnel surface.

Have fun!

Spencer Luster

LIGHT WORKS, LLC

USB3 fit for machine vision?

2011-12-14T22:38:00.000-05:00

USB cameras have never really been adopted for machine vision. Yes there are some products on the market – Lumenera seems to be leading the pack – but they’ve failed to break through. As I understand it, this is mainly due to the way USB tends to hog more CPU time than does FireWire or GigE, and because it’s difficult to have more than a single camera per PC.

Things seem different with USB3. This interface seems to be building a head of steam with vendors such as Point Grey, Ximea, and Basler all launching USB3 cameras.

What does that mean for we end-users? I think it means life just got a little more complicated. There I was, getting to grips with GigE, learning how to address the cameras (not as easy as the vendors would have you believe,) and now they expect me to try something else. And what does USB3 give that GigE doesn’t?

I’m not sure I know. Speeds are broadly comparable and cable lengths are plenty long enough for my industrial needs. One difference might be that USB3 supplies power, so I can save on a separate power supply. Yes, I could go PoE but that doesn’t really seem to have broken through, so camera choices are limited.

I don’t know about you, but I think I’ll hang back and watch how this all pans out.

Slicing logs

2011-12-13T22:09:00.000-05:00

Scanning technology is an interesting machine vision niche, and it’s one where there seems a preponderance of Canadian companies.

Why should this be?

Well Canada has a large lumber industry, and 3D scanning systems help maximize the utilization of every tree trunk. To see how this is done, take a look at the website of Hermary Opto Electronics. This British Columbia-based manufacturer of laser scanning systems has an interesting product family and some fascinating movies. These show graphically how a knowledge of the geometry, plus color or grayscale information, help maximize yield.

And if you want to improve your knowledge of scanning methods, there are some useful diagrams on their Scanner Technology page.

Punny title goes here

2011-12-12T22:12:00.002-05:00

“Free” is usually a good word. Especially when it precedes “machine vision software”. That’s why I clicked on a link that took me to the Ctrax website.

Ctrax is “an open-source, freely available, machine vision program for estimating the positions and orientations of many walking flies, maintaining their individual identities over long periods of time.”

Leaving aside the question of what is a walking fly – isn’t a fly that doesn’t, a “walk”? – I was surprised to learn that flies even have individual identities. I guess this just shows I lack the kind of education and brains needed to get in to CalTech, where the Ctrex development team are based.

But on a more serious note, it strikes me that this software tool may have other applications. I can for example see it as a low cost video analytics tool for anyone who wants to study the ebb and flow of people through a doorway. Or perhaps there’s a killer app in manufacturing that I just don’t have the imagination to see.

Wherever it gets used, there’s no doubt that those CalTech folks are a pretty smart bunch. One might even say, there’s no flies on them! (You may groan now.)

Getting in front of the camera

2011-12-11T22:20:00.001-05:00

Most of us in machine vision are engineers of some description, and that means generally we prefer to be behind the camera rather than in front of it. But if you’re in a Sales or Marketing type of role, at some point Andy Wilson or one of his people is going to point a camera at you and invite you to talk about your products.

That’s a great opportunity to promote both yourself and your company or employer – if you do it well. Sadly though, few engineers are natural performers, so we, the viewers, get to watch a lot of awkward shuffling and hear a lot mumbling, stumbling and “um’s” and “er’s”.

I don’t mean to be Simon Cowell-harsh - this is intended as constructive criticism. If you are in a role that requires you to present, get some coaching. And let me also mention that oftentimes it’s the people who think they are great presenters that need the most coaching.

Hopefully you’ll take my advice on-board, but then where do you go for feedback? Spouse? Co-workers? Well those are good, but what might be better is to find a semi-pro through the Toastmasters organization. No, they don’t propose toasts, and neither do they eat grilled bread, not in public anyway. What they do is help people become better presenters.

Seriously, please get some coaching before the next round of Vision System Design show videos.

Testing parameter changes

2011-12-08T22:32:00.000-05:00

Sometimes it’s necessary to a setting or two on your vision system. Maybe a threshold needs raising to cope with a change in the incoming material, or perhaps a new mold produces a slightly different dimension. Whatever the reason, there you are on the line, tweaking away, but how do you know the long term effect of the change you’re making?

Exeter Engineering, maker of potato grading equipment (and other things too,) has an interesting idea. Their Accu Vision Sorter has a “Last 500” function. This lets an engineer evaluate the effect that a parameter change would have had on the last 500 items inspected.

It’s an interesting idea, though I assume it also means storing the last 500 images, which may not be a trivial undertaking. It may however be a great way to increase both flexibility and robustness.

For details you’ll need to watch their video of the sorter. The inspection part starts at about 6 minutes in. You’ll find there are some other interesting points too, like a color selection tool to “show” the system what needs to be detected. There’s also mention of polarized filters on the color cameras. This last point intrigues me; I’m not sure what polarization would achieve in this application.

So, interesting ideas from looking at potatoes. Whatever next?

LED lighting – not as stable as we’re told?

2011-12-07T22:10:00.001-05:00

It’s probably ten years since LED’s started to make real inroads in the machine vision world. Up until that point we’d all used fluorescent and halogen lighting, but LED’s, so we were assured, were stable and long-lasting.

Turns out, that’s not strictly accurate. Yes, LED’s make a great light source, but they are not perfect. More specifically, heat is their great enemy. As an LED warms up its output drops. Not a lot, but perhaps enough to alter a critical measurement.

What’s the answer? Well CCS has a few tips on their website, under the heading of “Skillful use of LED lights.” Take a look and pick up a few ideas for your next project.

What Can Machine Vision Do For Me? A Management Guide

2011-12-06T22:21:00.002-05:00

I believe that machine vision, (taken to mean the industrial application of computer vision techniques,) has three roles within a manufacturing operation. It can be used to replace human visual inspection, it can be used for process control, and it can be a sensor that enables other automation. Let’s run through each of those points.

Replacing human visual inspection

Despite 30 years of lean manufacturing many factories still employ people to sort good product from bad. This is a task that machine vision can perform with far greater consistency. Software doesn’t get bored, it doesn’t get distracted, it doesn’t need regular breaks and it doesn’t alter the standard depending on what mood it’s in.

On the other hand, software struggles to deal with any deviation the programmer failed to think of. For example, imagine the blue in a printed label changes shade slightly. An inspector may spot this but will decide that the product is fit for purpose and will let it go. Machine vision however may send it all to the reject bin.

Inspection automation sounds great – lower costs and improved customer protection – but it’s really hard to make work

Process control

Machine vision is very good at presence/absence discrimination, so it makes a good error-proofing tool. Yes it may be more expensive than a dedicated poke-yoke, but because it’s reconfigurable it may be cheaper in the long run. It’s also good at monitoring processes and detecting problems, like a broken drill, and changes, like when ink is running out. Best of all, these types of vision applications tend to be reasonably quick and inexpensive to implement.

Enabling automation

One of the headaches in automation, especially of part handling, is that it can be really expensive to fixture product for a robot to pick it up. But by giving the robot eyes it can cope with variation in part location. In fact vision guidance can eliminate the need to have a person orient parts for the robot – which leads to some frankly rather bizarre automation installations – so it saves both capital (the dedicated fixturing,) and revenue (the part sorter.)

This type of vision is complex and expensive to implement, but the number of people with such expertise is growing and there are more implementations to take a look at. Machine vision is definitely changing the economics of factory automation.

There we have it: machine vision can cut costs and prevent defects. Just think bigger than automating the work the end-of-line inspectors do.

Do we need to explain the steak?

2011-12-05T22:07:00.001-05:00

Non-American readers will be puzzled by that title, so let me explain. There’s a saying here about selling the sizzle, not the steak. It means to tell the customer about the wonderful benefits from buying the steak rather than just showing him the meat. In short, I’m asking if it’s still necessary to start every management presentation with a definition of machine vision?

I raise this point because back on September 14^th I promised you a paper titled, “What is Machine Vision? A Management Guide.” I sat down to write it and realized that managers in manufacturing organizations probably know what machine vision is. They may not know of its strengths and weaknesses, nor will they be able to summarize the benefits of deploying such technology, but they do grasp that, in its crudest form, a machine vision system is a digital camera coupled with software that analyzes the images and returns data and conclusions.

So “What is Machine Vision? A Management Guide” has been replaced with “What Can Machine Vision Do For Me? A Management Guide” and if anyone disagrees with my logic I would love to hear why.

Surprising camera trends

2011-12-04T22:26:00.002-05:00

The latest survey of trends in machine vision and industrial cameras is out, and I find some of the conclusions surprising.

Specifically, camera and sensor manufacturer Framos, who carried out the survey, concluded:

Almost half of all cameras in use are color
There is an increased trend towards analogue
Users and manufacturers share a clear preference for lower frame rates.

There were other findings too – download the survey and read it for yourself – but those listed above ran counter to my expectations. I rarely see a need for color, I much prefer a digital interface, and in many applications frame rate is a concern. Am I out of step or does this tell us something about the survey methodology?

I suspect it comes down to who was contacted and how the questions were worded. Speaking as a user, my camera purchases are driven by the applications I’m working on. Color adds complexity to the task, so I only use it if it’s absolutely essential. A similar argument applies to higher frame rates: I’ll only pay the premium if my application needs it. And as for analogue … the twentieth century called and wants its cameras back.

Machine vision acquisitions news

2011-12-01T22:23:00.000-05:00

PPT Vision, despite being a pioneer in the machine vision field, has been struggling for a long time. It’s probably good news then that a White Knight, in the form of Datalogic S.p.A. of Italy has ridden to their rescue.

According to the press release on the Datalogic website, they are paying $5.2m for a business with revenues of $6m. That ratio – 86% of revenue – tells us that PPT was not doing well, despite their market reputation, an enviable patent portfolio, and a recently rejuvenated product line-up.

But what does this move tell us about the state of the machine vision industry?

Well to answer that. Let’s take a look first at Datalogic. This is a company with 2010 revenues of €393m (around $515m), making them bigger than Cognex, that is concentrated in code marking and scanning, with related ventures in machine vision (using rebranded products, I believe,) and vision-based retail loss prevention (Evolution Robotics.)

The PPT acquisition comes hot on the heels of Datalogic’s deal to buy code printing and reading company Accu-Sort for $135m (147% of revenues.) (Press release here.) Clearly then, as the Italians say in the PPT announcement, they see “interesting growth prospects” in machine vision.

Implications?

Well, Datalogic are now going head-to-head with Microscan in the code reading and machine vision field. It also means that Cognex is no longer the undisputed heavyweight champion, though their profitability and cash reserves are second-to-none.

Will this prompt a further urge to merge? In the machine vision camera world we’ve already seen AVT parent Augusta snapping up manufacturers with complementary product portfolios, and then there was Dalsa’s acquisition by Teledyne, so I think the answer is yes, expect more mergers and acquisitions.

As for me, I’m going to start a vision company and then sell out to one of these big players.

“Make it Right with Cognex Vision”

2011-11-30T22:26:00.000-05:00

No I haven’t started accepting advertising dollars, not yet anyway, (though that day may come,) hence the quotation marks around the title of this post. It’s a line from the song “Make it Right” performed by Ultrasonic Rock Orchestra.

Looking like a cross between Queen and some early ‘80’s New Romantics band (maybe the Human League?) URO perform a tribute to the wonders of Cognex Vision. It’s so … what’s the word I’m looking for … entertaining, that I had to reproduce it here for your viewing pleasure.

Now I’m more convinced than ever than Cognex needs to find a good use for all that cash they have stashed in the bank!

Gesture recognition

2011-11-29T22:23:00.002-05:00

eyeSight Mobile Technologies is a company doing interesting things with machine vision. They’re using the camera integral to almost every mobile gizmo on the market as an input device. The video on their website shows how a simple sweep of the hand can be used to turn an electronic page, adjust volume, or select from on-screen icons.

What intrigues me is that they’re doing this without 3D. They use a single camera and ambient lighting: there’s no IR projector as with the Kinect, and no time-of-flight hardware either. So how does it work? No details are provided, although I suspect there’s some kind of feature extraction – the hand – and perhaps some optical flow analysis to determine how the hand is moving. But I may be simplifying things too much. A quick Google threw up a link to this research on vision-based hand tracking.

The eyeSight video, naturally enough, makes the system look very robust. I’d be interested to know just how much precision is needed with the gestures – I can imagine a significant training time, just as there was with the old handwriting recognition systems. Neither am I convinced by all the applications shown. For in-car use I think voice recognition is a better way to go. In my oh-so-humble opinion, the killer apps will be those where messy hands need to be kept away from keyboards and mice. The medical field springs to mind, and perhaps manufacturing, especially in places where gloves have to be worn.

In short, this looks like a technology with a future, and judging by the funding they’ve received, I am not alone in this view. Now, who will be first to introduce a shop-floor version?

A robust vision system

2011-11-28T22:50:00.000-05:00

“Vision systems don’t work.”

How many times have you been told that? More than a couple, I’ll wager. And sadly, it’s too often true. You set the system up, check it’s making the correct decisions and walk away. A few weeks later, (maybe sooner,) you’re back because something changed.

This is the big problem: machine vision is totally repeatable, even if something changes. That’s a problem, because human inspectors will adapt almost instantly to acceptable variations, but machine vision? It’s so dumb it will just kick out all your good product, or worse, pass your bad product.

What’s to be done then? I could write a book about this – and perhaps I will – but for now have a look at “Ensuring Reliability in Vision Systems,” published in the November 2011 Assembly magazine.

This runs through the main issues in building a robust vision system: assess process variation, write a detailed specification, resist pressure to add additional inspections, get the lighting right, throw plenty of pixels at the task and so on. These are all good points and should be taken on-board.

But, (you knew there was a “but” coming, didn’t you?) I can’t resist nit-picking, so let me mention a couple of points.

First, there’s a big old typo: “Dumb cameras require an external image sensor, which initiates image capture.” This makes little sense to me: the sensor is always inside the camera, whether smart or dumb. Are they talking about triggering?

Second, Rick Roszkowski of Cognex says, “Always bring enough pixels to the party,” before going on to suggest a 10:1 rule – 10 pixels to span the tolerance band. Those aren’t bad rules of thumb, but:

Be aware that more pixels equals lower frame rates and more processing time – it’s a tradeoff.
I think that 10:1 rule is derived from the old metrology rule-of-thumb. I don’t disagree but I would dearly like to see some math applied to the issue of how many pixels are needed to produce a sufficiently repeatable measurement.

So, with those two little quibbles out the way, let me say again, “Ensuring Reliability in Vision Systems,” is a good read. Take a look.

Open wide

2011-11-27T22:36:00.000-05:00

One of the highlights of VISION, that took place in Germany a couple of weeks back is the presentation of the VISION Award. While the prize money of €5,000 is unlikely to provide much of a return on the investment, the publicity is surely worth a great deal.

This year the award went to The Austrian Institute of Technology for a 3D dental scanner. This is a hand-held device not much bigger than a toothbrush and containing two cameras and a blue LED. The idea is that a dentist, (or hygienist, I suppose,) uses the device to acquire 3D data of a patient’s teeth.

Details are rather sparse although optics.org does tell us the system includes “a miniaturized light projection system”. I imagine that means it’s projecting a series of parallel lines onto the teeth. Two cameras ensure there’s minimal shadowing. What I am curious about those is how the data is stitched together. I’m thinking there must be some kind of image to image registration algorithm.

The scanner is being commercialized by A.Tron3D of Austria, whose catchy slogan is “Yes we scan” but sadly they provide no info on the inner workings of the device. Presumably dentists don’t care how it processes the data, only that it does.

Now, open wide and say “Ahhhh.”

How to evaluate machine vision lights

2011-11-22T22:41:00.000-05:00

Go to a vision show and you’ll find any number of vendors showing off LED lighting. They’ll have red lights, blue lights, green lights, (not so many IR lights though – I wonder why?) and all sorts of geometries: backlights, darkfield ring lights, line lights and so on. Then, as you get into discussion with the sales guy you’ll learn that prices range from the somewhat expensive to the Oh-My-God-that’s-more-than-the-profit-we-reported-last-year. So how do you decide what represents good value?

The main criteria should be how well the light illuminates the object or area your vision system will be looking at. The problem I have though, is that without specialized equipment it’s virtually impossible to quantify uniformity of light distribution, or indeed, the spectral distribution.

So here are my guidelines:

Hold the light in your hand, (with it turned on – duh!) If it feels really hot, that’s a bad sign. Heat is the enemy of LED’s.

Examine the mechanical assembly. Is it rigid? Does it have good mounting points? Will it endure years of vibration? Can it be cleaned easily?

Study the LED placement. How uniform is the pitch? The height? Are any pointing off-axis?

While it’s difficult to assess uniformity, I prefer to see more small LED’s rather than a few honkin’ big ones. This tends to reduce the incidence of hot-spots.

And finally, value. While it’s a somewhat sweeping assumption, I tend to believe power consumption correlates with light output. So work out the price per watt. What this might show is that the differential between the cheap and the expensive light is not as great as it appears at first.

Obviously, none of this is as good as objective measurement of light output. But unless you’re going to buy a whole lab of test gear, I suggest these checks will help in determining which LED lights offer the best value for money.

Last, a quick note to my readers outside the US: this Thursday is Thanksgiving, so I'll be taking a break for a few days. Check back after the weekend!

VGR primer

2011-11-21T22:28:00.002-05:00

Go to any automation show and you’ll see robots guided by machine vision. They make it look so easy, although you know it’s not. But how do you go about getting up to speed with vision guidance?

Well you start by reading “The gift of sight: A four-step approach to setting up a machine vision system for robot-based manufacturing” This straightforward article, published on automationmag.com, November 17^th, 2011, discusses the main points to consider, in jargon-free English.

A particular strength of the article is how it covers robot-to-vision calibration. This is something I don’t think I’ve seen covered anywhere else, yet it’s essential (along with lighting,) to a successful implementation.

Read and learn.

Outlook for the machine vision industry

2011-11-20T22:17:00.002-05:00

With all the negativity the US media likes to force upon us (nattering nabobs, anyone?) it’s good to hear some more objective views. The latest cautiously optimistic report comes from optics.org in an article somewhat misleadingly titled, “Machine vision groups spark Asian expansion”.

Based on comments from Dr. Munkelt of MVTec and supported by data, (imagine that!) this tells us that growth should continue – albeit at a lower rate than previously – throughout next year.

That’s good news, but I do see an issue (or an opportunity): I think growth is going to be rather incremental, based on engineers using vision as a tool to solve manufacturing problems, unless there’s some technology breakthrough. In my humble opinion, what we need for accelerated growth is something that makes machine vision more robust and very much easier to deploy. Increased resolutions and higher speeds are all useful and will make those niche applications feasible, but if we could have a “magic” system with better-than-human capabilities I think the vision market would explode.

What are you working on?

Application tip for IR lighting

2011-11-17T22:33:00.000-05:00

Spotted an interesting little application story on the Vision Systems Designs website a few days back, “Smart cameras solve a pressing problem,” (Andy Wilson, November 1^st, 2011.)

This relates to the use of two Dalsa BOA smart cameras for checking that a die is clear before it closes. As the story relates, closing the die with slivers of stray metal in the way can cost a lot of money.

This seems like a pretty run-of-the-mill machine vision application but there are a couple of interesting aspects. Number one is that it’s the first time I’ve read about the BOA cameras actually being put to work. Number two is the use of IR lighting.

I don’t wish to steal Andy’s thunder, (or his ad revenue,) by giving you all the details here, but trust me, the application needed infrared illumination.

Now here are two tips for when you use IR. First, remember that silicon has low quantum efficiency at wavelengths much over 750nm (and goes to zero about 1,000nm,) so you’re going to need a lot of light to make the application work.

Second, most cameras have an IR cutout filter over the sensor. You’ll need to remove this in order to detect the IR light. This might seem obvious but I have found vision engineers complaining that an IR light isn’t working when all they had to do was remove the filter!

Just a couple of points to bear in mind.

Let’s hear it for the Swiss

2011-11-16T22:35:00.002-05:00

Today, for no other reason than that I think they have an interesting product family, I’d like to give a “shout out” to Leutron Vision. For some thirty years the good people of Leutron been making cameras in their factory near Zurich, yet few in the machine vision world know of them, so let’s change that.

There are two things I find interesting about Leutron.

Number one is that they were one of the pioneers in the field of PC cameras. Today they have Intel Atom-based cameras that will run a variety of vision software products, but they were getting in to this some four or five years ago.

Number two is their Simplon Software Developer Suite. This comes with programming libraries and a set of tools for camera control, but the cool thing is the code generation capability. In other words, use Simplon to set up your application, then turn it into code that actually runs the application. Best of all, and I’m 99% certain about this, Simplon is free.

So here’s the bottom line: if you’re interested in PC cameras, or just want to see what less mainstream camera vendors have to offer, check out Leutron Vision. I think you’ll find it interesting.

New cameras

2011-11-15T22:17:00.001-05:00

Over the last few days there’s been a torrent of new product announcements from German camera manufacturer AVT. Might this be connected with the Vision show held last week? How naive do you think I am?

Cynicism aside, these are some interesting products. Let’s start with the 6 Mp Prosilica GX2750. This uses a Dual GigE Vision® interface to deliver full-frame images from its CCD sensor at 19fps. Impressive numbers, to be sure. I have to say though, since I believe the pixels are just 4.5 microns and the cooling fins imply considerable heat generation, I am a little concerned about noise in the images.

Perhaps then I should look at their cooled camera, the AVT Bigeye with Active Cooling. Presumably resulting from the acquisition of VDS, this features a Peltier cooler to reduce noise. I have a feeling this is something we shall see more of as pixels shrink and resolutions grow.

Last off, (and I lack a good segue into this,) is the AVT Bonito. This is an impressively fast camera – 4Mp at 386fps – but what really intrigues me is the interface: CameraLink.

Okay, there’s nothing particularly earth shattering about CameraLink, but it’s new for AVT. They started out with FireWire, moving into GigE with the acquisition of Prosilica. And now CameraLink?

If other camera companies weren’t already paying close attention to AVT before, they should be now.

Financial review

2011-11-14T22:21:00.000-05:00

Today it’s the turn of Basler under the financial spotlight. Let’s see what we can deduce from their Q3 2011 results, published November 10^th.

In the commentary there’s a lot of talk comparing the first three quarters of 2011 with the same period in ’10. That’s hardly surprising because those numbers look good: revenues, gross margin, earnings per share, cash flow all improved. But third quarter comparisons don’t look quite so attractive: Revenues down 16%, earnings per share down 13%. Cash flow however was very strong, up 81%.

So what’s behind these numbers? Allow me to make three observations:

Headcount is up somewhere around 10%. At the same time, R&D and administration costs have increased.
The Solutions business seems to be struggling. Costs have been cut to match the decline in revenues but I have to wonder how much longer Basler want to be in this business.
Surveillance cameras are a bright spot, sales having doubled from Q3 2010.

What does this mean for the machine vision industry as a whole?

Clearly, it’s hard to make money in solutions. Selling hardware seems more lucrative. Also, it’s a competitive, product-led industry, hence the need to pump more resources into R&D.

And is there an industry-wide downturn? Basler say there has been a “cyclical downturn.” I’m not convinced though. I don’t recall the Cognex numbers suggesting a dip. I didn’t get to Vision 2011 this year but I’m curious what the mood was. Any readers care to comment?

High-speed imaging

2011-11-13T22:19:00.000-05:00

Since I’m “the guy who knows about cameras,” on occasion I’m asked for advice on high-speed video recording systems. Of course, these have little to do with machine vision, but they do produce some really cool images and movies, so here’s one I’d like to share.

One tip for your viewing pleasure: watch it full screen.

Out of respect for the originator, I should mention that I believe this comes from Fastec Imaging, so if high-speed is of interest, check out their website.

Familiar with LPR?

2011-11-10T22:20:00.002-05:00

If you answered “no” to that question, allow me to explain. LPR stands for License Plate Recognition and it’s the technology underlying the speed cameras that are proliferating throughout much of the world.

LPR is essentially an optical character recognition (OCR) application, but one where the whole image has to be searched and where the environment is pretty much unconstrained. As in all areas of machine vision, the technology is advancing all the time, so I thought readers might enjoy this peek at the state-of-the-art, courtesy of The Consumerist (“No Car Will Escape This “Future Now” Speed Cam,” November 1^st, 2011.) The video is particularly impressive.

If you want to get right to the source you need to visit Peak Gain Systems Inc, although frankly I found The Consumerist piece more informative.

Challenging the paradigm

2011-11-09T22:25:00.000-05:00

Every new entrant to the wonderful world of industrial machine vision is told the same thing: it all begins with lighting. Get the lighting right and everything will be straightforward. And historically speaking, that’s been true. Consistent illumination that creates contrast in the features you want to see makes it possible to apply standard vision tools, principally edge detection and blob analysis.

Now, think back to those two video analytics posts I made previously, Does lighting matter? and In-car video analytics. What do they have in common? No control over the lighting.

It was Mitch Rohde, CEO of Quantum Signal, who forced me to think about this. Back in the day when crude presence-absence systems first emerged, lighting was critical. But today we have so much number-crunching capacity available that perhaps we should think differently. Perhaps we should learn from our video analytics colleagues and extract information from video streams without worrying too much about the lighting and optics.

Can we disregard them completely? No, but should we continue investing time and effort in optimizing light direction, intensity and wavelength when maybe raw computing power can do it all for us? Perhaps it’s time to rethink this paradigm.

In-car video analytics

2011-11-08T22:25:00.000-05:00

I’ve been learning about the new E-Class Mercedes. No, not shopping, just daydreaming. And I’ve been very impressed by its safety technology. There’s even radar-based cruise control to stop you running into the car in front while checking email on your Blackberry. But is radar the best way to do this kind of proximity sensing?

Michigan-based Quantum Signal is putting much effort into automotive video analytics technology. On their website they show off Lane Departure Warning, Forward Collision Warning, and my personal favorite, Traffic Sign Recognition. (There is a video of each one.)

These are all impressive feats; I look forward to getting them on my car and avoiding tickets for speeding or failing to stop at a Stop sign, but what I want you to consider is how they do it. These systems, and the retail systems I covered previously, operate in what are unconstrained environments. To be more specific, there is no control over the lighting yet the systems are able to extract information from video.

Have I got you thinking?

Does lighting matter?

2011-11-07T22:25:00.000-05:00

The key to machine vision success is lighting, or so the experts tell us. But what if the experts are wrong? What if lighting is irrelevant to the success or failure of our applications?

Heresy? Perhaps, but I recently had the opportunity to listen to an image processing guru who argues our whole machine vision paradigm is wrong. And I think he may have a point.

Let’s start with video analytics – the business of extracting information from the torrent of data that is a video stream. Today I’m going to introduce you to video analytics in retail. My next post will cover the same technology in automotive applications, and then I’ll attempt to challenge our assumptions about how to “do” machine vision.

Intrigued? Then keep reading.

Lighthaus is a leader in retail video analytics. Their products will track people in a store and provide managers, layout planners and so on with information about which areas get the highest traffic, how long they loiter, and so on. They claim, and I will admit to finding this a little hard to swallow, that their technology can even estimate the age and gender of visitors to a store. That’s something I struggle with, despite my advanced processing capabilities and years of accumulated knowledge, so I imagine there’s a pretty high margin of error.

If you spend some time on the Lighthaus website you’ll find a number of movies showing their systems in operation. It’s impressive stuff, and despite my skepticism regarding some of their claims, I think you’d have to agree that it works. In my next post we’ll take a look at the same technology in the automotive world.

Multi-sensor product verification

2011-11-06T22:12:00.002-05:00

Much as I love machine vision, I recognize there are some things it can’t do. It can’t for instance, tell what capsules are in a pill bottle. Yes, it might be able to give you their size, shape and color, but is that enough?

When it comes to pharmaceuticals, I’m not one to take chances, which is why I enjoyed “Machine Vision and Raman Spectroscopy” by Kenneth W Chapman, posted on The Imaging Source blog September 9^th, 2011.

This describes the addition of Raman technology to a machine vision system so as to determine exactly what’s in those pill bottles. This might seem a somewhat narrow niche, but I don’t think it takes too much imagination to see how the concept could be applied to more industrial problems. How about checking on the composition of a surface coating, for example?

Machine vision can’t do it all, so let’s keep our eyes open for complementary sensor technologies that can help us avoid saying “No.”

Cognex responds

2011-11-03T22:35:00.000-04:00

If you read my previous post, Cognex results for Q3 2011, you'll recall that I was a little scathing about how they spend their R&D dollars.

Well it seems my post caught the eye of someone in Natick, because an anonymous commenter, (Dr. Bob perhaps?) wrote to say that the fine men and women of Cognex have not been "sitting on their haunches" and have in fact been very busy improving the Dataman code readers.

Actually I did know that, but I decided not to let a few facts get in the way of a good poke. And, to be honest, while code-reading is technically machine vision, it's not the kind of machine vision I'm really interested in. What I'd like to see are smart cameras with liquid lenses and even smarter defect detection algorithmns. After all, Keyence are working hard on those kinds of innovations, and they clearly have Cognex in their sights.

Those Natick haunches are clearly not being sat upon, but at the same time, they should look to their laurels. (How's that for a mixed metaphor?) Didn't some business guru say every business should be trying to obsolete itself?

Cognex results for Q3 2011

2011-11-02T22:27:00.000-04:00

I have three ways of gauging the health of the machine vision industry: I measure how long it takes a salesman to return my call, (longer equals busier,) I count how many times a week a salesman wants to stop by to show me a new product, (more means less busy,) and I look at the Cognex financials.

Right now the first two measures say business is good, so let’s see if the Cognex third quarter results confirm that finding. (Here’s a link to their pdf.)

Some interesting points caught my eye. First, to summarize, sales are up over Q3 2010 and year-to-date 2011 versus same point in 2010. But income (or profit for any Brits reading – I know you’re out there,) is down. Obviously then, expenses have increased, but which ones?

Well Dr. Bob has two explanations: the sales team is growing, and they’re putting more money into R&D.

To the first point, one might expect that a bigger salesforce would equate to more sales – in fact shouldn’t it be a somewhat linear relationship? – but I suppose there’s a lag while the new hires ramp up. And to the increased R&D spend – well that ought to be good, but I must confess to wondering what all those boffins in Natick are actually doing. I mean, we don’t seem to be awash in new product announcements.

I also noticed that yet again, Cognex can find nothing better to do with their piles of cash than return it to shareholders and buy back their own stock. Neither of these is exactly bad, but they do seem be indicative of a business that has matured and is no longer pouring forth innovations.

So what should they be doing? Well allow me to make two points. First, perhaps they want to look closely at their innovation processes. Maybe they need a skunk works to come up with the Next Big Thing. And second, perhaps a few strategic acquisitions? After all, there are quite a few innovative little machine vision start ups out there, and some might have some good ideas.

There again, Cognex is doing very nicely as it is, so maybe they’re following the “if it isn’t broke, don’t fix it,” strategy.

Advice on lens selection

2011-11-01T22:11:00.000-04:00

If I was to categorize the machine vision questions I get, number two on the lens would be optics. Every camera system needs a lens, but lens nomenclature makes no sense to the typical vision user. In an effort to help, periodically I direct readers to the best lens calculator package I know of – MachVis, buried deep on the Qioptiq website – but sometimes that’s not enough. What’s needed is a simple optics primer.

It seems the folks at Basler have reached the same conclusion, because they’ve recently put a two page white paper titled “Optics Recommendation” on their website. It’s not the definitive guide, but it might help you on your way.

And what do I get most questions about? That would be lighting.

Scandinavian machine vision lighting

2011-10-31T22:08:00.000-04:00

I’ve only been to Scandinavia once: the people were warm and friendly, the beer was fantastic, the weather was a bit ho-hum, but what really hit me were the prices! That is an expensive part of the world!

I wonder if that applies to lighting products from LATAB?

As with everything that comes out of Sweden, (Volvo, Abba, pickled herring – or is that Norway?) they look fantastic, but I have a feeling they will be a wee bit expensive. That said, the machine vision lighting business is pretty competitive, so perhaps I’m being unduly harsh. If you want to do some comparison shopping you could do worse than talk to these guys.

Let me know what you think.

Benefits of machine vision inspection – porosity detection

2011-10-30T22:29:00.002-04:00

Casting processes tend to produce porosity – tiny bubbles in the metal – and these are uncovered when the metal is machined. Porosity on sealing surfaces is a problem because the product may fail in service, so to prevent customer complaints, machined products are often inspected. As this is usually done by human inspectors it represents an opportunity for machine vision.

Most vision projects start out by trying to achieve the same level of defect detection as the human inspector, and that’s where things get interesting. If you look at the work instructions or quality standards you’ll see that they usually call out a minimum size for porosity of about 0.300mm.

Does this mean that the sealing surfaces don’t leak if the porosity is smaller? Not at all. The minimum size is based on what an inspector can see. 0.300mm is about the size of a speck of pepper, and since porosity looks black against a white surface, if it’s this size or bigger it’s usually detectable. Any smaller and there’s a high probability it will be missed.

But if the point of inspection is to ensure the customer gets only good product, shouldn’t the minimum size be dictated by the product and not the eyes of the inspector?

If larger porosity – up to 0.500mm is acceptable, then perhaps good product is being scrapped. After all, the inspector doesn’t measure the porosity, they just reject if any is seen. In this scenario machine vision will actually reduce waste.

Conversely, if the product still leaks at 0.300mm then perhaps the standard needs to be tighter. Machine vision can do that, so in this scenario it will lead to improved product performance.

Now let me throw a question at you. Most quality standards for porosity are defined by area, but pores have depth too. A pore with a 2D area of 0.250mm might still create a leak if it’s 1.000mm deep. What should we do about that?

Something for the weekend

2011-10-27T22:44:00.000-04:00

It’s not often machine vision gives us chaps an excuse to look at attractive young women, so I thought I’d save this post for Friday and the weekend. And I should mention that this comes courtesy of gadget blog gizmag.

You’ll recall that the driving force behind the Kinect was Primesense. Well Primesense have now moved into the tricky business of 3D body scanning. Partnering with Bodymetrics, they have produced a machine that maps the human body and advises on clothes that fit.

Okay, I realize that few of us scruffy male engineers are going to rush over to London, England to get scanned, but it does give us a great reason to watch the technology in action.

Strangely enough, no portly, middle-aged male engineers were featured in the “The science behind the fit” video, which I heartily endorse, purely for its overview of the 3D technology of course.

Learning from others

2011-10-25T22:19:00.000-04:00

There’s always something to be learnt by looking at how someone else solved a problem, and a great example of that comes from the October edition of Vision Systems Design. Under the delightfully alliterative heading of “Detecting Defects in Dinnerware” (October 1^st, 2011,) comes a highly detailed description of a sophisticated inspection system.

I don’t want to steal author Dan Milkie’s thunder, (or VSD’s advertising revenue,) by repeating the story here, but let me explain why you should take a look.

The article describes an inspection system developed by Coleman Technologies of Newton Square, Pennsylvania, that performs both 3D and linescan imaging. There’s lots of detail on how both are done, including an excellent description of the image processing performed. If you have any interest at all in surface flaw detection algorithms, this should be a must-read. If you visit Coleman’s website you’ll also be able to view a movie of the system running in the development shop.

Learn from what other people have done.

Laser lines without the laser

2011-10-24T22:17:00.000-04:00

Over the last few weeks there’s been quite a bit of discussion on these pages about how to deal with speckle. (And thank you to all those who have taken the time to comment.)

“Speckle” is an issue with laser lines, since it complicates the business of finding the line center. If you’re doing laser triangulation-based 3D machine vision, you’ll know just what I mean.

Several commenters have reminded me that an alternative to the laser does exist. Smart Vision Lights have a family of LED Structured Light Pattern Projectors, as indeed do Opto-Engineering. I’ve not tried the projector from Smart Vision Lights, but I did have an opportunity recently to get “hands-on” with the Opto product. This did a fine job of projecting a uniform line, but I have to say that it lacked the intensity of a laser. In a 3D application that might mean making a trade-off between speed and precision, but it’s definitely something worth exploring.

These projectors have been around for a while, but Opto-Engineering has now taken them further with the addition of Schiempflug optics. These “straighten out” the distortion caused by projecting a pattern of lines from an angle, and if you’re interested in learning more, there’s a great video on Opto’s YouTube channel.

Finding out who my friends are

2011-10-23T22:25:00.002-04:00

Regular readers will have noticed my list of “Friends worth checking out” over to the right of this page. This serves two purposes.

First, it’s part of my mission of providing useful machine vision information, so I link to those who offer relatively uncommercial discussion of vision topics.

And second, it’s my feeble attempt at a bit of search engine optimization. The idea is that when folks link to me, I reciprocate. Call it mutual back-scratching, if you like.

Well I’ve noticed that some of those cherished in-bound links aren’t there any more. So guess what? I’m not sending any traffic their way either. I’m just not feeling the love.

On the other hand, when people with something useful to say about machine vision link to me, I am very happy to link back, which is why the “Friends” list changes. If you’d like to be my friend, and perhaps attract some of my readers, email me at visionmaster05@yahoo.com and we’ll try to sort something out.

Multi-sensor inspection application

2011-10-20T22:22:00.000-04:00

From the Basler “Success Stories” files comes “Ultrasound Diagnostic Railway Car with Basler scout GigE Cameras for Synchronized Visual Inspection “ (it open up as pdf.) This describes a system where ultrasound is used to detect cracks in rails. What makes it challenging is not so much that this is done at speeds of 60km/h, and in all conditions, but that rails actually have holes in places. This means the ultrasound system is going to give indications of flaws when none are present.

To deal with this, a camera system, synchronized with the ultrasound, captures images of the rails. This makes it possible to review a suspect area of track and decide if further investigation is needed. Obviously, it’s more complicated than how I’ve just described things, so if you’re interested, take a look at the note on the Basler website.

Verifying datamatrix codes

2011-10-19T22:26:00.000-04:00

An engineer called in a panic. His customer had rejected a delivery because they couldn’t read the datamatrix code, yet when he ran the offending parts through his system he got good reads every time.

This is an all too common problem and it boils down to the difference between “reading” and “verifying”. “Reading” happens when you point the scanner at the code and it outputs a character string. The problem is that the customer almost certainly does this at a different working distance and with different lighting, so what works under your ideal conditions may fail in his suboptimal environment.

It was to overcome this that “verifying” was introduced. Verification is done under standardized conditions and with the objective of assessing the quality of the marking. I sent the panicked engineer a link to a useful article explaining the difference, and if you’re interested I suggest you look at it too. (I would cut ‘n’ paste the relevant paragraphs but was too intimidated by the threatening copyright notice.)

If you click the link above you’ll find that the article references various standards, most notably ISO 15415. This provides a standard for code verification but unfortunately there is a problem. As explained in “New standard improves verification of Data Matrix codes” (Test & Measurement World, February 2007,) the ISO standard is geared to black codes printed on white paper. It doesn’t work quite so well with codes laser-marked on metal surfaces. And that’s where AS9132 comes in to play.

But putting aside details of which standard to follow, the point remains: the engineer and his customer needed to agree a common way to assess code quality. Without this, arguments over readability, or lack thereof, devolve into shouting contests that no one will ever win.

Why Scorpion? Thor replies

2011-10-18T22:30:00.000-04:00

Back on October 6th I asked why I should consider Scorpion for robot guidance rather than any of the “packaged” solutions available from robot vendors and their partners. Thor Vollset of Scorpion was kind enough to post a couple of comments in reply, and I thought readers would be interested in seeing his response.

So to my question, Please tell us what Scorpion provides that I can’t get from my robot supplier’s existing partner, Thor replied as follows:

The first unique feature is: The system does not require ANY changes / tuning to pick ALL the parts produced in the factory.

We can copy the cell and run the same machine vision in all the robot cells in the factory. This is what 3D machine vision is all about.

These requirements have been pushed upon us by the AutoCast project manager and he has provided resources and time to implement this and succeed.

When the project started in 2009 we knew that we had to lift our 3D platform to be able to 3D pick from a bin based on the information from 3D Stereo Vision cameras. We have felt all the challenges where we get frustrated about the fact that the corners of the bin is so damned dark. It is hard to locate every object all the time - the contrast is so low - it cannot be possible. To overcome all these issues we have improved our sub-pixel accuracy both in 2D and 3D, we have invented the easiest to using multi-core processing to meet the cycle-time demand of 5 seconds - we have succeed on speed using quad-core cpus - with the new penta-core we have software that will easy run faster using the 2x6 = 12 cores on these systems. Even more has been invented and more information is found at the TordivelBlog.

The video on the blog is boring by purpose. But is shows how the system will be used in the factory and the video is a big milestone in the project - we can empty the bin without any problems without ever failing over and over again due to the extreme accuracy.

The system can pick object in a bin 1200 x 800 and 600 mm height. The accuracy is a lot better than the customer needs sub mm in x,y and z.

We can handle any number of objects with x and y rotation up to +- 30 degrees. This is more than the robot can handle!

We are proud to have implemented a robust system that will improve the competitiveness of the Norwegian casting industry and reach the goals of the project. The platform is also suitable for other kinds of 3D Picking in all kind of industries.
*
*
*
After my technical focused reply I saw that I did not answer your question: Why choose Scorpion and not the solution that ABB, Fanuc, Kuka and other vendors provide. This question is not simple but I will try to answer.

Since we at Tordivel started making robot vision ten years ago a lot has changed. 2D was ruling and we got a lot of attention just stating we provided robot vision. We learned that making a prototype is easy, but delivering up to the expectation with robot vision is very hard. The picture has changed - most vendors do supply 2D vision and a lot of people supply 3D Robot Vision. I have been trying to promote machine vision to these companies with some luck but often with no luck.

Why is this? - I believe we will succeed because we care and understand the issues and exceptional values of accuracy and robustness. We can fix it when the system stops. Scorpion 3D never stops! Scorpion 3D Stinger is so good! When I talk to the big vendors with the standard robot vision - nobody likes robot vision, nobody talk about the fact that systems are stopping numerous time over a week and the loss of productivity. The automation people normally get scared when you talk about the great potential and the tremendous opportunities and value in 3D Robot Vision.

We have cared about our systems for more than ten years releasing a better software every Friday - now build 496, we do not accept to deliver a system that stops several times per day. We deliver great systems. We know what to do to get it working. Our systems are ten times as accurate and lot better than these standard solutions. We can solve application that the standard systems cannot even dream to solve.

The devil lies in the details we care about!

We deliver stereo vision that never fails with Scorpion 3D Stinger and Scorpion Vision.

Keyence goes linescan

2011-10-17T22:44:00.002-04:00

As far as I know, the last machine vision company that attempted to make linescan imaging accessible to the regular user was DVT, and that wasn’t hugely successful, so I think it’s pretty big news that with their new XG-8000 product, Keyence is moving into linescan imaging.

The XG-8000 is actually the vision processor unit, but it will take a linescan camera, two if you also buy the expansion module. As with all Keyence vision products, the cameras are proprietary, meaning that you can’t plug in your Dalsa or Basler cameras. But there are some interesting features.

In terms of resolution, they’re offering 2k, 4k and 8k sensors, all of which are CMOS. The smaller sensors take a C-mount lens while the 8k takes a special M40 lens. Going to a C-mount format makes the cameras smaller than most linescans, and also permits a shorter working distance, both of which might be useful when space is tight.

No word on pricing - in fact at the time of writing the XG-8000 isn’t even on their website – but my guess would be around $15 to $18k. But I am guessing! That makes it expensive for a single camera configuration, but if you need two cameras the economics become more attractive. I imagine that the Keyence linescan cameras would be pretty much plug-and-play, which could be quite a time saver, and it also saves you from having to buy a framegrabber.

In terms of software and vision tools, the last Keyence system I played with was quite impressive, so I have no reason to think this will be any different. One other benefit I see is that while most linescan systems need some programming expertise, Keyence doesn’t.

In short, if you want to implement a linescan system quickly, and you’re not a skilled C++ or VB programmer, this might be a good product to consider.

Even More on Despeckling

2011-10-16T22:28:00.000-04:00

Here’s another technical note and handy tip from Spencer Luster of LightWorks, this time in relation to some earlier discussion in these pages regarding the issue of laser speckle.

Just when you thought it was safe to turn on the lights, here are even more notes about despeckling laser sources!

Speckle is a function of two kinds of source coherence: 1) Temporal (or spectral) and 2) Spatial. The first refers to the purity of the frequency/wavelength (color) of the light. Spatial coherence refers to how well all the waves of light line up with each other—are they all in phase. The more coherent the light source, the bigger the potential speckle problem.

Temporal (spectral) and Spatial Coherence.

Same frequency/wavelength and all in the same phase. Very Speckly!

If you don't have temporal coherence you can't have spatial coherence, but the opposite isn't true. All the identical frequency light waves could occupy a huge number of different phase positions. You would have perfect temporal coherence, and near-zero spatial coherence. What's more, you'd have little if any speckle problem!

No Temporal Coherence and therefore also no Spatial Coherence. No Speckle.

Temporal Coherence, but no Spatial Coherence. Little or no speckle.

The following techniques can help reduce spatial coherence. These are just a start. I'm sure you'll think of others if you're interested.

Ping pong balls: Drill two 1/8" diameter holes in a ping pong ball, 90 degrees apart. Fire your laser into one hole. After multiple diffuse reflections inside the ball, the light exiting the second hole will have greatly reduced spatial coherence. If you use it as a light source, the image will have very little speckle. (Some will recognize this as a poor man's integrating sphere. It also wastes a lot of light.)

Twisty light pipe: Take a 1/8" to 1/4" diameter acrylic or PETG rod, 2 to 3 ft. long. Polish the ends. Use a very hot water bath, or a heat gun to soften the rod. Carefully bend it into a spiral shape of 2 to 3" radius. Pump laser light in one end. After many reflections inside the spiral, during which many different optical paths are traversed, the output will have reduced spatial coherence. I've built such systems that work well, and are reasonably efficient.

You can pull the same trick with large diameter liquid filled light guides, or even multiple fiber optic strands of different lengths. The point is to significantly alter the path lengths of lots of portions of the light. ("Normal" small core fiber alone usually doesn't let enough path length differences build up.)

The above and similar techniques don't produce collimated beams, but you do get very spectrally pure sources with little or no speckle. Some applications call for this, such as thin film inspection or critical polarization measurement.

Spencer Luster

LIGHT WORKS, LLC

F-stops, resolution and depth-of-field

2011-10-13T22:36:00.000-04:00

One of the most baffling aspects of machine vision technology is finding the “ideal” lens settings. Unless you’re serious about photography, (and I am not,) terms like “f#” and “diffraction limited” might as well be in Mandarin. This is why I lap up any and every article I can find on optics, and is the reason I read “Into the Depths,” (Vision Systems Design, September 2011,) several times.

Written by the very knowledgeable Greg Hollows of Edmund Optics, this article attempts to clarify the relationship between the size of the aperture, (indicated by f#,) the depth-of-field, and the resolution of an imaging system. Unfortunately though, I still found it too complicated and had to resort to Wikipedia.

And what did Wikipedia tell me? Well as part of MV4U’s mission to simplify machine vision for users, I shall summarize and then direct you to some useful sources.

f# - this tells you how open or closed the lens aperture is. Much confusion comes from the fact that a small f# means the aperture is wide open while a big number means it’s very small. Thus at f# of 1.2 the aperture is wide open and much light can pass through, while at f# of 16 the aperture has been closed down to a pinhole and hardly any light can get through.

When light passes through a pinhole it gets bent, or diffracted, a lot. This doesn’t happen so much with a bigger diameter hole. Diffraction is a bad thing because it reduces the resolution of the imaging system. In simple terms, the more diffraction that takes place, the more blurry the image. So to get a good image, open the aperture wide, (or as a photographer would say, use a small f#.)

But here comes a problem: when the aperture is open wide the depth-of-field is very shallow. In other words, the image is only in focus over a very small change in working distance – maybe just a few millimeters. Close the aperture down (increasing the f#,) and the range over which the object stays in focus increases.

So how does the eager vision engineer balance these tradeoffs? Well from Wikipedia I ended up at “Lens Diffraction & Photography” on the cambridgeincolour.com website. (Be sure to read both pages.) This has two excellent diffraction limit calculators; one finds the circle of confusion, the other estimates the f# at which resolution will start to be compromised. I found that playing with these in conjunction with the Greg Hollows’ article help make his explanations clear.

The other useful calculator I found was for estimating depth-of-field. Again, this is useful for trying things out and seeing what effect they have.

Lenses are mysterious things for most of us engineers. Hopefully these links and my explanation will help bring it all into focus. (Loud groans please!)

Vision systems and PLC’s

2011-10-12T22:44:00.000-04:00

Vision systems rarely run in isolation. Usually they are interfaced with a PLC that handles part presentation and disposition of the inspected items, and in my experience, this is where a lot of integration issues arise.

At the heart of the problem, I believe, is the issue that machine vision people want to deal with optics, light, image acquisition and processing, and many of the breed have no “controls” expertise. On the other side of the divide sits the controls guru. He can make a PLC sing and dance, but vision systems are not his “thing”. The result is that, far too often, “What we’ve got here is a failure to communicate,” as Paul Newman said in Cool Hand Luke.

To help address this, the September 2011 edition of Vision Systems Design has an excellent article titled, “Programmable Logic Keeps Machine-Vision Systems Moving”. This describes the basic principles of PLC – vision system communication before going on to propose an additional concept, the “Clear-to-Send” PLC output.

The idea is to improve handshaking between PLC and vision system by having the PLC tell vision when it’s ready to receive data. (I think that’s the gist of it.) This is an excellent idea, except I’m not sure it’s new.

In my on-the-job training years ago I was introduced to the idea of a Busy signal. As far as I can tell, Clear-to-Send is the same thing as NOT Busy. Now it may be that I’m missing some subtleties of inverting the logic, so if I am, please send me a comment or email to explain.

I’m not trying to be Mister Negative; I think this is a great idea, and a great article, (Andy – we need more like this.) I’m just suggesting that perhaps those skilled in the art of PLC-vision system integration already know.

Linux machine vision?

2011-10-11T22:23:00.000-04:00

I look at Linux the same way I do electric cars: it all sounds very attractive but I’m really not sure I want to stray that far from the tried and true. I’m not alone in this because to my knowledge there’s only one Linux vision product on the market, and that’s Ned Lecky’s Voyant Vision.

But that is changing with the release of Common Vision Box 2011 MultiOS. This latest version of the venerable CVB package will, according to parent company Stemmer, run on 32 and 64 bit Windows as well as 64 bit Linux. That seems to me like a bold move intended to future-proof CVB against the shift to 64 bit operating systems currently underway.

My 64 bit experience is very limited but it’s my understanding that developing on 32 bit for a 64 bit OS, or vice versa, is tricky to say the least. Obviously, if you’ve gone 64 bit Windows as a way to access more RAM, the Win 32 limitation has to present a problem, but I suspect there’s more to it than that.

Nevertheless, I like the direction Stemmer are going in liberating the vision software from the constraints of the OS. It’s almost enough to make me try Linux.

Useful vision tool from Cognex

2011-10-10T22:25:00.000-04:00

Sometimes a vision system that’s run trouble-free for months, or even years, will start rejecting good parts. You go out on the floor to take a look and can’t see any obvious problem yet clearly something has changed. What do you do?

My advice has been to acquire a master image when the system was first signed-off. Then, when problems arise pull out that master and compare it with the latest images. If light output has diminished, if the trigger is out or if the part under inspection is actually different, that will be readily apparent. But no one seems to do this.

Well if I’ve understood the new “TestRun” tool from Cognex correctly, that’s pretty much what it does. It appears, (though the press release could use a little added clarity,) that this allows a check on current image quality, comparing it with a baseline and flagging any changes.

What an excellent idea!

Machine vision entering the public consciousness?

2011-10-09T18:33:00.000-04:00

If you’re tired of having to explain to your wife/husband, kids, mother, or golf buddy what machine vision is, here are a couple of links to help.

First, from Singularity Hub comes an excellent graphic explanation under the heading of “Robot See, Robot Do…Robot Understand. Machine Vision Is Changing The World – Here Are The Details.”

And second, The Daily WTF has a short story, (yes that’s right, I said story, and I think it’s fiction but I can’t be certain,) titled “Caught.” This concerns out of control machine vision and robots, (or is it the users that are out of control? You decide.)

I find it fascinating, and even quite exciting, (yes, my life is that dull,) that machine vision is finally becoming a mainstream technology. Clearly the Kinect, plus all the code-reading software now in our smartphones, is encouraging a growing curiosity about how images can be interpreted by software.

Are we vision people on the verge of being cool?

Why Scorpion?

2011-10-06T22:10:00.000-04:00

Staying with the vision-guided-robotics (VGR) theme of my last post, the marketing machine that is Thor Vollset recently drew my attention to his work on 3D bin picking. This was in support of the Autocast project, which is something I’ve been following closely.

In a blog posting on September 13^th Thor shows a video of a robot emptying a bin of castings. (“This is Autocast 3D Bin Picking for the Casting Industry.) We could nit-pick about the extent to which the castings were in a known and ordered arrangement for the video, but in the main it’s quite impressive.

But here’s my question: the last time I visited the Robots & Vision Show in Chicago every robot company had vision guidance. It seems that all of them, or at least the majors like my corporate robotics partner, is already in bed with a vision company. This means that if I need robotic bin picking I call up ABB, Fanuc or whoever and they quote me a turnkey package. I’m not sure why I should disregard their standard offering and instead go with Scorpion.

Thor, I know you’re a regular reader so perhaps you’d like to comment. (Make it a good one and perhaps you’ll get your own guest blogger spot!) Please tell us what Scorpion provides that I can’t get from my robot supplier’s existing partner.

A tip for vision-guided robotics

2011-10-05T22:40:00.000-04:00

It seems like everyone is doing vision guidance: using vision to guide a robot to pick up a randomly oriented part. Most often, the next step is then to insert the picked part into some kind of assembly or packaging, and this is where problems can start to emerge.

Often, when the robot gripper clamps or sucks down onto the part it will move slightly. This can cause problems if the next step is a precision insertion-type operation, so here’s a tip.

What you need is a registration step. Have the robot show the part to another camera. The computer can then use this image to calculate the positions of key features with respect to the gripper. These new coordinates can then guide the assembly step.

Now this isn’t my idea. I’ve seen several integrators do this kind of step, but one of the most interesting applications was covered in a case study from Xigent Automation Systems of Ohio. This is described in an Assembly Mag article, “Automation Profiles: Integrator Sheds Light on Assembly Challenge” published August 24^th, 2011 but for better photos you need to look at the Robotic case study on Xigent’s website.

Overview of 3D vision applications

2011-10-04T20:32:00.000-04:00

Are you familiar with the game of “Blemish” often played by insecure managers? Its where, no matter how good your work is, the boss finds some tiny flaw to pick away at. It has a pretty demotivating effect and I hate it, so I’m going to play it now.

Over on the AIA website (www.machinevisiononline.org,) you’ll find an interesting article by Winn Hardin titled “Through the Looking Glass: Differences Between Large and Small 3D Vision Applications” (September 21^st, 2011.) It’s a little too tilted towards the Gocator from LMI for my tastes, but otherwise it’s a good reminder that 3D covers the gamut from microns to meters.

I do however have a problem with one sentence. Now this could be my lack of understanding, and if so please point out my error, but I am perplexed. Winn says “Structured light solutions … are well-suited to high-precision 3D imaging applications, as well as for isolating hard-to-image features such as the bottom of a screw hole where the pitch is so steep that it’s difficult to accommodate the different views from a two-camera stereovision solution.”

My understanding of structured light is that it requires an angle between camera and light (typically it’s a sheet of light, like a line from a laser.) So if the light is coming in at 45^o, doesn’t that make it well-nigh impossible to check the depth of a screw hole?

If I’m wrong I would like to know, because I do have a few applications that would benefit from measuring deeply recessed features.

Looking the gift horse in the mouth

2011-10-03T22:13:00.002-04:00

I was taught long ago that if something seems good it’s often wise not to ask too many questions. Today though I want to break that rule by asking why machine vision sales are growing so strongly.

We know the vision business is healthy – a look back at “Business and economic news” will give you some data – but I’m curious why?

An article on the ControlDesign website offers three suggestions. It’s all down to increased processor power, with a little help from the increases in camera resolutions and framerates, plus a small boost from the ease-of-use of GigE cameras.

I don’t disagree with those being the more significant of the technology trends, but I think there’s something deeper going on. Yes, businesses, especially manufacturing companies, are willing to invest in reducing the Cost of Quality, and that’s part of it, but I think the biggest driver is the number of people out there with machine vision knowhow.

It’s you dear reader. You are the driver, because you are able to talk intelligently about what machine vision can do. You’re a disciple, spreading the message that machine vision actually works. And as others get that message they too become disciples, evangelizing for vision.

Stick at it, because you’re doing a great job.

Don’t use a sledgehammer to crack a nut

2011-10-02T22:11:00.002-04:00

We’re machine vision professionals! No job is too big or too small for our extensive image capture and processing experience! But yes, sometimes we do get carried away and, dare I say it, over-engineer solutions. Sometimes all a task needs is the right sensor.

But what sensor, and how to use it?

Well allow me to point you to a YouTube channel – emxinc’s Channel – that shows how color and UV sensors can be used to solve a variety of presence/absence type applications. There are over a dozen short videos, and they’re not getting a lot of views. (Perhaps they need a dog that barks “I wuv Emx.”)

I do realize none of these will knock your socks off. Oscar material they are not, but I suspect that at the same time you’ll be like me and think these are rather neat ideas. After all, simple is usually best.

Machine vision and our civil liberties

2011-09-29T22:19:00.000-04:00

Two threads I keep returning to in this blog – I write it so I’ll decide what goes in to it – are the increasing use of vision for vehicle safety, and the way in which camera technology may be about to intrude on our personal freedoms.

Today I’m able to merge those two themes into one, by sending you to “Speed cameras out, Big Brother in?” on the website of “Engineering & Technology”. This article explores how various authorities want to slow down drivers who exceed posted speed limits, given that speed cameras, which spread like a nasty rash all over the UK, are now gradually being withdrawn.

A number of possibilities are suggested, and much of the discussion is about building on technologies already in place, such as lane departure warning and collision detection systems. (I suppose that should be detection of collision probability exceeding some threshold as it’s pretty easy to tell when a collision has happened.)

Now I’m all in favor of reducing traffic fatalities, but I am also a big fan of individual freedom, so I find myself on the fence on this issue. What I think is clear though is that we – by which I mean society – need much more of a debate about how far we are willing to have our activities captured by camera, and how we want those images used.

Rant over. I’ll get back to the techie stuff soon, I promise.

Matrox smart camera application

2011-09-28T22:33:00.000-04:00

“A Complicated Vision,” (ControlDesign.com, May 2011,) is the kind of machine vision application story I like to read. It concerns the use of multiple Matrox Iris cameras to verify the correct assembly of engines at an unnamed manufacturer.

Of particular interest are the reasons for going with the Iris, (no, I’m not going to tell you – click the link,) and the level of detail provided about the particular inspections performed. I wish we could have learned more about the lighting – that’s often left out of application stories – and an insight into the actual measured benefits would have been good too, but we can’t have everything, can we? Still, even without those details, it’s an interesting read.

Why we need a debate about facial recognition

2011-09-27T22:29:00.000-04:00

If you’re on Facebook at some point you’ve probably been “tagged” in a photo, and being knowledgeable about computer vision, you understand that was done by facial recognition software.

This worries me.

The problem I have is that cameras are spreading like a plague. It’s not enough that everyone except me carries an 8 megapixel camera with them all the time, it’s all the security cams, also known as CCTV. There are just far too many such video systems for humans to watch all the feeds, so they have to be analyzed by software. And, as discussed in this fascinating BBC article about some research at Carnegie Mellon University, facial recognition software is rapidly moving in to the security realm.

Big Brother really is watching us.

But there is another point of view. One of the people interviewed by the BBC notes that two hundred years ago pretty much all of us lived in communities where everyone knew who we were. There was no urban anonymity when we lived in towns of just a few hundred people, and maybe that was a good thing. Having grown up in a small community myself, I know full-well that fear of someone telling your parents what you’ve been doing helps keep you on the straight-and-narrow!

My worry though is that without some debate about how far we want facial recognition to go, we’ll wake up one day to find it’s happened, and at that point it will be too late. So let’s talk about it. Let’s tell our friends and families where this technology is taking us, and let’s see if they think that’s a good thing or not.

If we don’t at least talk about it we risk waking up in a world we might not like.

Machine vision in the greenhouse

2011-09-26T22:48:00.000-04:00

It’s been a while since my last agri-vision post, so allow me to make up for that by sharing details of a fascinating system. This is the “Scanalyzer” from German company, LemnaTec.

The Scanalyzer is a vision system that acquires images of plants grown in research greenhouses, as when scientists are trying to optimize fertilizers or select plants for breeding based on particular characteristics. Software then analyzes the images, extracting the required information. It take sonly a little machine vision experience to spot some of the challenges in doing this: even if you control and standardize the environment, each plant will be different, so how do you apply a common algorithm?

LemnaTec have some great videos explaining the system and some of their algorithms on YouTube. Here’s one, but for others I suggest you to go to YouTube and search for LemnaTec2008.

Lasers, speckle and sub-pixel interpolation

2011-09-25T22:20:00.000-04:00

One task that benefits from sub-pixel interpolation is finding the midpoint of a laser line, typically if you’re doing “sheet-of-light” laser triangulation 3D-type work, (where a line is projected onto a surface from an angle and viewed from directly above, or the other way around.)

The problem I find though is that the laser line doesn’t look as good to the camera as it does to my eye. This is due to the speckle inherent in collimated laser light. One way to deal with speckle is to open up the lens aperture, creating a saturated line. But, if you then threshold the line, the edges are uneven, which means the calculated midpoint also wanders about. (Yes there are other approaches to finding the midpoint of the line, but I’m not going to go into them right now.)

This is discussed in an interesting machine vision tutorial written by one Jose Dolz, who is a rather smart computer vision expert from Spain. Under the heading “Subpixel Detection” Jose addresses morphological approaches to the laser line problem. If this is something you’re struggling with, take a look at his tutorial.

One downside of morphology is that it sucks up processing time, but there is an alternative: eliminate the speckle.

I touched on this back in May 2009 in “Dealing with laser speckle” which makes three suggestions, but more recently I’ve come across a fourth option: a “laser speckle reducer.” These clever gadgets are sold by Edmund Optics and they’re not cheap, but depending on how well they clean up your line, it might be money well spent. I’ve no experience with them so I can’t comment, but if you’ve used them I’d like to hear how that worked out.

Do you believe in sub-pixels?

2011-09-22T22:19:00.000-04:00

Measurement is a complicated subject, and nowhere does it seem more complex than in the world of machine vision. Allow me, if I may, to outline the problem.

Imagine you wish to measure a box. You grab a steel rule, only to find that the smallest unit on the scale is an inch. Thus you can say the box measures between 10 and 11 inches along its side, but you cannot give a more accurate number.

Flip the rule over and you find it is graduated in centimeters (I know the centimeter is not an ISO unit, but I’m old-school.) This allows you to say the box measures between 26 and 27 cm. That’s better than before (because centimeters are smaller than inches,) but it’s still not very precise.

Let us first recognize that this might be good enough. If the purpose of the measurement is to cut a piece of wrapping paper that will cover the box, an uncertainty of +/- 0.5 cm is probably enough. No need to measure to three decimal places.

But what if you need a more precise measure? Well by eye you can subdivide the gap between 26 and 27 cm. It’s easy to estimate the midpoint, and you can probably get pretty close on the quarters. What you’re doing is interpolating between 26 and 27, and that’s what sub-pixel interpolation does. It looks at the gray values of a series of pixels and mathematically estimates where an edge would be.

I’m not a fan of this approach because I’ve never seen a vision system produce two identical images. There’s always some variation in pixel values, even if only because of noise in the system, but other machine vision folks, some with far more expertise than me, will argue that sub-pixel interpolation is a valuable tool.

Readers who’ve been with me a long time might recall that I addressed this back in January 2010 under “A case study on sub-pixel interpolation”. At the time I admitted that I’m a skeptic but I did refer you to a paper by Den Dawson of Dalsa. The link I gave then still works, so if you’re curious, take a look.

If you’re looking for a job…

2011-09-21T22:36:00.000-04:00

There seems no shortage of machine vision jobs these days, which means that salaries do appear to be creeping up, but if you’re in need of a new opportunity where should you look?

I advise people to start with the LinkedIn machine vision groups: I’m seeing an increasing number of positions being advertised in those, and there are several recruiter-members, so even if you don’t see something that grabs you, they’ll no doubt be pleased to hear from you.

Dice used to have a number of machine vision jobs, though for some reason I’m not seeing so many there. (Perhaps they’ve all migrated to LinkedIn.)

However, I was recently sent a link to the 50 best engineering job sites, and that does appear to be exactly what it says, so you might want to spend a little time there.

Good luck!

Machine vision out of its depth?

2011-09-20T22:34:00.000-04:00

Regulars to these pages know that I like to see machine vision doing good deeds, and here’s a great example of just that.

A company in France has developed a system that monitors swimming pools to detect possible drowning. The system, which goes by the name of Poseidon, uses multiple cameras mounted both above and under the water to keep track of bathers. Proprietary software analyzes the images and sounds an alarm when indications of drowning are detected.

Now I’m not sure what those indications would be, but there are some hints on the website as to the nature of the algorithms used. These imply there’s some motion tracking going on, with deviations from “normal” being flagged as alarms. I would imagine that a body at rest on the bottom for more than a few seconds would trigger such an alarm, as would a swimmer who stops moving forward and begins thrashing about wildly. There’s also a reference to stereo vision, presumably to derive some information about the location of each swimmer.

So far it appears the Poseidon system has only been sold to operators of large public pools – where it has saved many lives – but wouldn’t it be good to see it a standard feature of every backyard swimming pool? Think how many lives could be saved by a low cost drowning detection system. There’s a challenge for all you vision entrepreneurs.

Half the story

2011-09-19T22:32:00.002-04:00

Using machine vision to mount wheels on cars sounds like a pretty interesting application. Logically, there must be at least three parts to the task: determine the rotation of the studs or threaded holes on the wheel hub, and do the same on the wheel itself. Then pick up a wheel, rotate until it’s in angular alignment with hub, place and insert bolts or screw on nuts.

But that’s not what “Smart Camera Helps the Wheels Go ’Round and ’Round” in Vision & Sensors (August 22^nd, 2011,) describes. This article tells us how Matrox Iris smart cameras are used to check hole location on a wheel and to verify the correct wheel has been picked. We are also told that the robot mounts the wheel while the car moves past but, unless my reading skills are slipping (and they may be,) I could find no reference to how the wheel is aligned to allow insertion of studs. Yet this has to happen for the wheel to be secured.

A visit to the integrator’s web site, IBG Automation GmbH, (part of Goeke Group,) was only marginally more helpful. The details are in German but as I read it, this is a 3D application where the cameras are used to check angular orientation of both wheel and hub, guiding the robot appropriately.

I would like to know much more about this. For example, how is the lighting configured, and does the robot need depth perception (Z-axis) to know how far to move when mounting the wheel?

It’s a very interesting application but unfortunately Vision & Sensors is only telling us part of the story.

An argument for in-house machine vision expertise

2011-09-18T22:24:00.000-04:00

A vision system is not like a machine tool. Sure, there’s a frame or structure and often there’s motion, but whereas a machine is engineered to be robust, machine vision is like a delicate child that needs frequent care.

The issue is that vision systems, much like some of our young, are very sensitive to changes. The biggest headaches often come from subtle variations in the object being inspected – a switch in paper vendor can result in different reflections and an altered image, to give just one example – though small shifts in the operating environment can be just as troublesome. Imagine what will happen when a conveyor belt is changed for an “improved” version, or perhaps as dust builds up on a fixture (not a food industry example, obviously!)

If there’s no in-house vision expert able to trot out onto the factory floor and nurse the system, what happens? Well it means a call to the vision system distributor or integrator and a production stoppage until they can visit.

“But wait,” you cry, “Why not connect over the internet?”

Well that’s a great idea but even if your IT group allows outside parties access to your network, it doesn’t always work in practice. How will they refocus a lens that was accidentally knocked during a 5S cleaning exercise? How will they spot that the ink on the label is a slightly different shade of blue?

Sometimes there’s just no substitute for being there. So either you ensure your support is close at hand, or you brace yourself for extended downtime while the tech books a flight, or you have people on site who can troubleshoot these problems.

I prefer the latter, but it’s your factory so it’s up to you.

The best way to skin the cat?

2011-09-15T22:22:00.000-04:00

Sure, we know there are many ways to solve a given vision problem (“skin the cat”,) but is there ever a ‘best’ way? Well when it comes to web inspection I think Thor of Scorpion Vision has certainly hit on a smart idea.

Under the heading of “Scorpion 3D Surface Defect App” Thor describes a triangulation-based 3D approach to web inspection. There’s also a nice little animation to make his approach clear. What I like about this is how it eliminates the effects of varying surface reflectivity and inconsequential marking.

Of course, if it’s surface marking you’re after you might want to go with a more conventional web approach. A short post on the Vision Systems Design website alerted me to one company offering just such. This is Slovenia-based Wise Technology, who has a system called “Wise Scan.”

Technical details are rather sparse, but as best I can tell it’s a fairly conventional web system. That doesn’t mean it’s no good; it’s probably perfectly good for many applications (and perhaps quite price-competitive too,) but I feel that Thor is adding something extra.

Now, how about combining the 3D with regular web inspection? That would yield a mass of data about the surface, perhaps increasing robustness too. Any venture capital folks interested in backing me?

A question about motherhood and apple pie

2011-09-14T22:32:00.000-04:00

“Motherhood and apple pie” is a phrase Americans use to describe things that few would disagree with. It’s often used as a way of saying that someone is stating what my English colleague calls “the bleedin’ obvious.” And it applies to an article in the latest Vision & Sensors newsletter from Quality magazine, “Vision & Sensors: Machine Vision Tips: Advances in Vision Sensing” (September 1^st, 2011.) Now that’s an attention-grabbing title, so I dived right in, and boy was I disappointed.

I was expecting to read about, well, advances in vision sensing, but what I got was a primer suitable for someone very senior in an organization, like my bosses boss. All the article told me was that cameras make great sensors, they can do lots of neat stuff and I should be using them in production. (I’m half expecting said bosses boss to email it to me with a note saying, “We should look into this.”)

But am I being too hasty? Is there still a need for management guides to machine vision? Do people still want to read vision primers?

Google says that each month 74,000 people search on the phrase “What is machine vision”, so perhaps there is a demand. You know what this means don’t you? Look out for my N part series titled “What is Machine Vision? A Management Guide.”

Trends in machine vision cameras

2011-09-13T22:21:00.000-04:00

CameraLink provides the highest robustness and data rates, FireWire has been superseded by GigE, and USB is irrelevant for machine vision. Oh, and by the way, resolutions are going up.

Well that’s what I told an engineer who emailed me a few days ago asking if he should stick with FireWire cameras, but let’s take a moment to dive deeper.

Few in machine vision would argue that CameraLink, for now at least, provides the best overall performance. It can handle the highest data rates, (up to 6Gbps with 2 cables,) works over cable lengths as long as 10 meters, (enough for most industrial applications,) and never misses a frame, but it does have downsides. Chief among these is cost. CameraLink needs a framegrabber, which can easily cost $500 or more. It also needs special or custom cables – yet more cost. And since Power-Over-CameraLink (PoCL) doesn’t seem to have really caught on, perhaps because it’s limited to 4W, figure on running two or three cables out to the camera.

So what about the alternatives? Well let’s start with USB. Every computer has a USB port, so there’s no need for a framegrabber, making it seem attractive. But when you dig down into the details the problems start to emerge. The basic issue is that USB is something of a processor-hog, thanks to its employing a master-slave architecture. There are also bandwidth issues, meaning that a significant proportion of its 480Mbps is not available to transfer image data. For these reasons few manufacturers of industrial cameras have ventured into this standard. (Scientific and astronomical cameras are a different matter of course because they don’t need high data rates.) The maximum cable length of 5m can also be bothersome in many applications.

And then there’s FireWire, IEEE1394, to use its proper name. 1394 cameras, which need no framegrabber and take their power – up to 45W – from the single cable, come in two flavors, a) and b), the difference being speed. Roughly speaking, a) handles 400Mbps while b) hits 800Mbps. That’s more than enough for a 2Mp camera running at 5 fps, so what’s the problem?

Well first, cable length is limited to 4.5m, which can get to be a problem in even the simplest of industrial settings. Second, and I’m speaking from personal experience here, it’s just not totally reliable. I’m not up in all the details of error detection methodologies, but it’s my understanding that 1394 doesn’t have any. That means a frame can be dropped and you’ll never know. Now in many applications that might not be a problem, but if you need to be 100% confident in your system, think carefully before going this way.

That brings us to GigE, strictly speaking, GigE Vision. This uses plain old network cabling, though you should upgrade to the better shielded CAT 6, to transmit up to 1Gbps of image data as far as 100m. You will need to buy a special purpose card for your PC – the regular network card can probably not handle Jumbo Frames and in any case, you need it for internet and intranet access – but these are far less expensive than a CameraLink grabber. And if power cabling is a concern, Power-over-Ethernet (PoE) will deliver up to 13W – enough for most cameras.

Downsides? Well really, just the fact that you’re limited to 1Gbps, which is more than FireWire b) can offer, so it’s enough for run-of-the-mill applications.

As for resolutions going up, well you knew that already, didn’t you?

Business and economic news

2011-09-12T22:22:00.000-04:00

Two interesting items crossed my desk today. The first was a press release from the AIA titled “Machine Vision Market in North America Posts Healthy Results in Second Quarter 2011”. This good news came as something of a surprise, given all the doom and gloom we’ve been hearing. I suppose though that the outlook was brighter back in the April to June time period where sales were reported as 16% higher than in the same period in 2010. It may also be that some vendors had pushed through some fairly steep price increases, which would inflate the numbers if not the actual units moved.

The second item was an announcement from Basler of their half year results. (Find it under “Financial Reports.”

Basler earns less than a quarter of its revenue from North America, so this may reflect trends in different parts of the world, but the key points were that sales were up 22% and earnings up 50%. So if they saw only the same increase in North America as the rest of the industry it must be that other markets are really booming.

And to my point about price increases, Basler noted that they are making record numbers of cameras. So it’s not all from inflation; it seems their business is still growing quickly.

If you’re interested in comparing these data points with those from other companies, take a look back at “Business outlook for machine vision” which I posted on August 4th.

This is all good news, but there is unfortunately a cloud on the horizon. In their mid-year report Basler noted that, “The global economic upturn has lost momentum in the second quarter.” So is it time to panic?

I don’t think so. I believe we’ve seen the end of the rebound and we’re back to the steady growth our industry has been seeing for many years. The outlook, in my humble opinion, remains good.

Myths & Facts about GigE Vision

2011-09-11T22:26:00.000-04:00

The GigE interface is growing rapidly in popularity with machine vision professionals, but misconceptions about its abilities seem rampant. To address this I thought I’d write a post or two about the pros and cons of GigE, but then I found Scott Israel of 1stVision had beaten me to the punch.

So without further ado, here’s a link to Scott’s paper entitled, “GigE Cameras: Myths & Facts”.

Is Baumer the only company to have figured out PoE?

2011-09-08T22:36:00.000-04:00

Power-over-Ethernetseems like an obvious idea: make GigE a single cable camera interfacelike FireWire or USB by sending the juice down the same wire as theimage data. But a search for a PoE interface card has me rethinkingmy assumptions.

Letme take you through my thought process.

Inoticed recently that AVT now offer a PoE version of their Mantacameras.I haven’t used the Manta but it seems an attractiveprice/performance offering, so I thought I’d take a look. Now theManta comes in regular GigE and PoEflavors, and as I like the idea of not two but one cable, I decidedto search for a suitable PC card.

Tomy surprise, the only company I could find with such a product wasAdlink Technology. There may be others but Google doesn’t knowabout them and besides, it was the AdlinkGIE62+card that was recommended on the AVT website. This looks like a goodproduct but I like to comparison shop, so I searched the websites ofother PoE camera makers to see what they would recommend.

Thatwas how I ended up at Baumer.Now this German company has a fine range of cameras, but what reallygot my attention was “theGigE Vision complete solution”.

TheBaumer approach to GigE is to inject the power at some point betweenPC and camera. They’ll sell you a GigE Power Switch to do this, butin conjunction they’re pushing a “GigE Trigger Device”. Thisappears to be a micro plc that will sit in the electrical cabinet,counting encoder pulses and issuing trigger commands to camera andlights.

Thisstrikes me as an eminently sensible approach. It means you run asingle Ethernet cable from PC (presumably using an Intel Pro100card,) to the controls cabinet, and from there you run a singlecable, this time with electricity, to each camera. And as a bonus,you don’t need to program a plc to handle timing and triggeringissues.

Thisis why I’m thinking that only Baumer have the right paradigm forPoE.

Comments?

What skills are needed for a job in machine vision?

2011-09-07T22:33:00.000-04:00

In an effort to help job seekers, in “Machine vision – where the jobs are” I mentioned the two fields where I see recruitment activity taking place, and labeled them (perhaps unkindly,) as those who do and those that think. As a follow-on, I thought it might be useful to summarize what background and skills you need to get a job in those fields.

Let’s start with the thinkers – those who carry out research in machine vision. Companies like Cognex, Microsoft and Google seem to be hiring people for vision research but I think most of the opportunities are in universities. Landing one of these jobs will almost certainly require a PhD in machine vision, image processing, or a closely related field. On the plus side though, this is the way for vision people outside of the US to get a visa to work here.

I’m going to spend more time on the skills needed to “do” machine vision, because this is the area I know better. So here’s a list, in approximate priority order, of what I look for when I’m hiring new team members. (And no, I have no open positions right now. Sorry.)

This is what I like to see on a resume:

An excellent grasp of machine vision physics (lighting, optics, how a camera works – that kind of thing.) An interest in photography is a real plus because it means you really know depth of field, aperture settings and so on.
Strong problem solving skills, and I do mean solving rather than calling up a vendor and asking for a solution. You must be able to define the challenge and conceive and evaluate multiple solutions. Great tenacity and attention to detail will really help you out.
A good hands-on, action orientation. While thinking is good, I like to see people getting out in the workshop and trying ideas out. Don’t just talk about it, do it!
A variety of machine vision experience. This will sound harsh but if you’ve only used smart cameras from one vendor I think your experience is pretty limited. You should be familiar with multiple vision products, and preferably you’ll know PC vision too.
Good interpersonal skills. Good engineers tend not to be people people but if you work for me you’ll be expected to sell ideas as well as educate users in the details of machine vision. You’ll need to be able to look people in the eye and smile as you explain in non-technical terms why they must expect a proportion of false rejects.

You might notice that I’ve left out specific programming skills, controls and motion control expertise and electrical knowledge. It’s not that these are important – they are – but they’re meaningless without the points I’ve bulleted above. Okay, not meaningless, but they won’t get you a well paid machine vision job.

One last point, and this is aimed at those outside the US looking for a company to sponsor their visa application: it’s not going to happen. Sorry but no integration house is going to invest the money it takes to set someone up in the US. They’d rather pay top dollar for some home grown talent. So if you’re looking for machine vision to be your ticket to the land of the free, well aim for one of those ‘thinker” jobs I described above.

Machine vision – where the jobs are

2011-09-06T22:28:00.000-04:00

If you live in the US you’ll be well aware that there don’t seem to be enough jobs to go around. I believe things are similar in Europe ‘though I’m afraid to say I’m less well informed about South America and Europe. But one area that seems awash with jobs is machine vision. Everywhere I look I see companies seeking vision engineers, so this seems like a good time to explore where to look and the skills that are needed.

The first thing to say is that machine vision employment can be split into two broad categories. There are the doers, those who develop and implement working vision systems, principally for inspection, and then there are the thinkers. This group works more in the realm of theoretical and applied research, addressing topics more in the field of image processing – algorithm development, video analytics and so on.

So which category is doing all the hiring?

Well it seems the answer is both. If you go on to LinkedIn and search for machine vision jobs you’ll find a wide range of work, from hands-on integration to really blue sky/long grass, dreamy stuff. As you might expect, the latter tends to be located on the West Coast while the former is more evenly distributed across the country. That said, my impression is that Atlanta and Kansas City, MO seem to be hubs for machine vision integration. Atlanta might be a result of the presence of DVT back in the ‘90’s but Kansas City I’m at a loss to explain.

Now obviously I’m writing from a US perspective, so I’d like to hear back from you. Where do you see machine vision jobs, and where should readers go to look for them? Use the ‘Comment’ function or just send me an email.

The influence of heat on LED lighting

2011-09-05T22:26:00.000-04:00

LED lighting is pretty much ubiquitous in machine vision these days, and for good reason: LED’s last forever and they’re totally stable.

Well that’s what we’ve been told, but it’s not completely correct. Yes, LED lighting is superior to fluorescent in that it’s a monochromatic source that shows very little intensity reduction with age, but it is vulnerable to the effects of heat.

Perhaps the biggest issue with LED’s is the risk of thermal runaway – they draw more power as they get hotter until they go bang (well actually they just die, but you get my point.) This is why it’s so important to use a current driver with LED lighting, and it’s also advisable to provide a big heat sink. But there are a couple of other issues too.

As noted in LED’s Magazine, light output reduces as temperature rises – another good reason for a big-assed heat sink – but what I find really interesting is that this effect is related to the wavelength of the LED itself. Specifically, red LED’s are much more temperature sensitive than are the blue ones. An argument for using blue or green lighting rather than red?

A second issue, noted in the LEDs Magazine article and reiterated on The Bergquist Company website, is that the wavelength of the emitted light is influenced by temperature. In fact LEDs Magazine notes an increase of as much as 0.13nm per degree C. Okay, in a factory environment that might mean a shift of only 2nm, but that could be enough to reduce the number of photons getting through a bandpass filter to your CCD.

So what to do? Well don’t give up on LED lighting: it’s the best we’ve got. Just be conscious that temperature can be a problem. I suggest good heat sinks and airflow are the way to build robustness into your vision system.

An annoying application story

2011-09-01T22:33:00.000-04:00

I’ve just read a run-of-the-mill vision application story in Assembly magazine that really made me quite cross. Now usually I’m a fan of their vision articles, but “Vision System Enables Manufacturer to Achieve 100 Percent Quality Inspection” (July 28^th, 2011,) was disappointing.

It wasn’t that not a word was said about the lighting that enabled this Cognex InSight system to detect cracks in ceramic parts, though that would have been interesting. Rather, it was the statement that, “The vision system also eliminated the problem of good parts being detected as faulty.”

Don’t misunderstand me, reducing false rejects is good, eliminating them would be wonderful, if that was indeed what the integrator had done, but it wasn’t. The problem was that some parts were “discolored due to dripping oil.” So does that mean the vision lighting and algorithm were smart enough to handle such contamination? Well here’s how they did it: “Now, the assembly machine stops automatically once three faulty parts pass through the test system in succession.”

Excuse me but that hasn’t eliminated the problem of good parts being rejected at all. All it’s done is reduce the up-time or OEE of the equipment. I was hoping to learn about some clever new technique, perhaps involving a Pat tool from Cognex, that made the oil drips invisible, but no such luck.

I know everyone is short of time and it’s tempting to reproduce the press release without spending a lot of time on analysis, but please, let’s not repeat claims without thinking them through first.

A new approach to buffers

2011-08-31T22:20:00.000-04:00

Buffers are one of those obscure aspects of machine vision system development that you don’t need to know about until you transition from smart cameras to PC-based systems. And when you make that move you’ll probably be left scratching your head over the what, why and how.

Ned Lecky provides one of the best explanations of buffer management that I’ve come across on his VoyantVision web site (look under the “In-Place Execution” heading,) but even this is lacking, so I’m going to have a go myself.

A buffer is, to put it crudely, space where your image can be copied. But why would you want to do that? Well as Ned explains, this becomes necessary when you are going to change the value of a pixel, based on the value of its neighbors.

This doesn’t occur if you’re performing simple tasks like edge detection because while you look at pixel values, you don’t change them. It doesn’t apply to thresholding or binarization either because, although that operation changes the value of a pixel, it works on that pixel in isolation. But when it does happen is when you’re performing a convolution.

A what?

A convolution. Convolution is when you multiply the image with a second, much smaller image. There’s a great explanation of this on the HIPR2 website of Edinburgh University, which also explains that image processing tools such as Gaussian Smoothing and the Sobel operator use convolution.

The issue is that convolution changes the value of a pixel, based on the value of its neighbors. So, if you change a pixel value and put it back in the old image, when you convolve the adjacent pixel, that operation will use the new value of the last pixel processed and not its original value.

To get around this, when you’re processing an image, typically to improve it in some way, a buffer is a place to put the new pixels that result from your convolution. Unless you’re Ned Lecky, in which case you’ve worked out a really clever way to put the new pixel values back in the old location.

Buffers: you never knew you cared.

Advice on selecting a lens

2011-08-30T22:37:00.000-04:00

Choosing a lens depends on two things: the field of view you will be imaging and the size of the sensor in the camera you selected. (If you haven’t selected a camera, take a moment to read over “Advice on selecting a camera – how many pixels?” and “Advice on selecting a camera – sensor considerations”.)

The ratio of sensor size to field of view (which you might also think of as object size,) is the magnification required, but you’ll notice that lenses are rarely specified in terms of their magnification. More often, you’ll see focal length, which relates to magnification.

I’ll spare you all the mathematics (Wikipedia does a great job – just look up ‘focal length’) and cut to the chase by saying that the shorter the focal length, the more rays of light entering the lens are bent towards the center or axis of the lens. This means that a shorter focal length lens – like one of 8mm - bends light more than one of, say, 50mm. And the consequence of bending light more is that the lens ‘sees’ over a wider angle. Thus to image a small field of view that’s a long way off, use a long focal length lens.

Now given the working distance, (the distance from the front of the lens to the object you’re imaging,) and field of view, you could calculate the focal length, but why bother when other people have written software to do it for you? Just Google ‘machine vision lens calculator’, search my blog, or best of all, download the free MachVis lens software from Linos (now Qioptiq.)

A word of warning is in order though. The lens calculators are not 100% accurate. This stems from the manufacturing tolerances in both cameras and lenses. Plus, if you use the MachVis software its results are specific to their lenses. If you use lenses from another manufacturer distances will change slightly.

So how do you deal with this? Simple: ensure there’s a certain amount of adjustment in the height of your camera. That way you can vary the working distance until the image fills the sensor, with a border to accommodate optical distortion and light fall-off, then screw it down and put it to work.

Appreciating telecentricity

2011-08-29T22:31:00.000-04:00

Telecentric imaging, and lighting, are valuable tools for the machine vision professional, yet they seem poorly understood. To remedy this, over the last four weeks Spencer Luster of LightWorks has contributed a series of four articles taking an in-depth look at the subject. If you’ve been busy, here’s another chance to see … as the TV people like to say.

Telecentric Lenses with Collimated Light: Part I, Subtle Defect Detection

… in which Spencer shows us how to use a telecentric lens and a collimated backlight to highlight flaws in transparent materials.

Telecentric Lenses with Collimated Light: Part II, Alignment

Here Spencer discusses the importance of alignment and adjustment.

Telecentric Lenses with Collimated Light: Part III, Spot Size

Which should be bigger – the image spot size or the telecentric entrance pupil?

Telecentric Lenses with Collimated Light: Part IV, Collimated Dark

Doing clever things with a darkfield backlight.

Advice on selecting a camera – sensor considerations

2011-08-28T22:03:00.000-04:00

Number of pixels is, as we discussed previously, the number one criteria in selecting a camera, but there are some others to consider. Here a few bullet points you might want to take on board:

Frame rate. This matters because it drives the communications standard you pick. Slow applications can generally make do with FireWire. These days, for the midrange I like to use GigE, while for high speed imaging (and especially with linescan cameras,) I always go with CameraLink. In the future I may have to add HSLink to this list, but for now CameraLink handles pretty much anything I throw at it.
Pixel size. Small pixels give better fine resolution because the positional uncertainty is reduced, but they capture fewer photons per unit time than do larger pixels.
Spectral sensitivity. Maximize the number of photons you capture by matching the wavelength of the lighting to the peak sensitivity of the sensor. Interestingly, most sensors today seem to peak in the green yet we continue to use red LED illumination in most applications. Perhaps it’s time to go green?
Noise. Frankly I put this well down the list of parameters to consider because with good lighting it’s possible to minimize the effects of noise in most industrial applications. I will note though that if you are struggling for light, cranking up the gain or exposure will accentuate any noise in the image (which manifests as gray levels that vary from image to image.) Now if you’re doing astronomy or scientific imaging noise may well be a more important consideration.

Lastly, we should probably mention the great CCD versus CMOS debate. Old school vision guys will tell you that CCD’s always give better images, but I’m not sure that’s as true today as it was a decade ago. However, there are a couple of differentiators to be aware of.

First, CMOS sensors come with either a rolling or global shutter. This can have a huge impact if you’re imaging something in motion, so make sure you understand this point properly. Second, CMOS sensors have better windowing capability, meaning that if you can accept a reduced image size you can turn the frame rate up to some pretty astounding levels – it’s all driven by bits per second coming down the pipe from the camera.

So recapping, the imaging task will dictate the resolution of the sensor required. Once the number of pixels has been established it’s time to decide which of the other parameters matter. That will result in a specification, and from that point it’s time to get out the credit card and go shopping.

Advice on selecting a camera – how many pixels?

2011-08-25T22:25:00.000-04:00

You’re new to machine vision. You know you’ll need a camera and lens, but when you open the Edmund Optics catalog (or start surfing the web,) you’re overwhelmed by the options. So, in keeping with my mission of providing practical advice to machine vision users, here are a few tips:

The task dictates the hardware. Don’t buy a camera because you like the color, because there’s a pretty girl in the catalog or because it has more megapixels for the buck than any other camera. Start with the task.

Define the field of view – the area that you need to image. Don’t worry about fitting it to the proportions of the sensor (16:9, 4:3 or whatever,) just determine length and width, or diameter if that’s more appropriate. (Incidentally, I prefer to work in metric units – I just find it makes life easier.)

Now, a word of caution: Never plan on using every pixel in the sensor. This is because optical distortion and a fall-off in light intensity make the border pixels less reliable. To avoid using these I suggest adding 10% to the longest dimension you need to view, and calling that the field of view.

Now you need to determine the resolution required, in terms of pixels per millimeter. Okay, this can cause some head-scratching so let’s dive a little deeper.

The drivers for pixels per millimeter are either the smallest feature you need to detect, OR the required measurement resolution. If you need to check the presence of a screw in an assembly, then the smallest feature you need to find is that screw head. But how many pixels does it take to do that?

People have tried to apply some science to the business of how many pixels make a feature detectable but I prefer to rely on old-fashioned heuristics. (That’s “rule-of-thumb” to you!) The absolute minimum number of pixels needed to find a feature is a square measuring 3 pixels by 3. But, if it’s a circular feature you need to find this can result in only the center pixel fully covering the feature. The border pixels will have a grayscale value partway between that of the feature and the background, making detectability difficult. Thus 4 by 4 pixels is better and 5 by 5 better still.

Now, how about measurement resolution? Without going in to how an edge is detected, let’s just say that calculation of edge location depends on the gray levels of the surrounding pixels. And you can’t assume that the edge lines up neatly with the pixels; you’ll never have a white pixel adjacent to a black pixel – there will always be one or two shades of gray in between. So my heuristic for edge detection is that the uncertainty in the measurement will always be 3 pixels.

But before you apply that to the tolerance on the part print, remember this: you almost certainly have two edges to find in your image, so the total uncertainty is 6 pixels, not 3. And how is this related to the part tolerance? Well opinions vary, but I suggest that the measurement uncertainty should be no more than 10% of the tolerance on the dimension. So if you’re measuring a feature that should be 25mm +/- 0.6mm, (I picked those numbers to make the arithmetic easier,) the uncertainty should be 0.060mm, meaning that you’ll want each pixel to span 0.010mm.

And with that number, you can calculate how many pixels are needed to image your field of view.

A stake in the ground

2011-08-24T22:33:00.000-04:00

The next time you replace human inspection with a vision system, do you think Production will complain about the false reject rate?

I think so. In fact I’ll bet a slap-up dinner for two that they will. “The system’s kicking out too many parts,” someone will bleat, “And it’s costing us money.”

Well since you know this will happen, why not take action beforehand to head it off? Here’s what I suggest: gather some data on the accuracy of your human inspectors. Juran tells us they’ll be around 80% accurate, though I suspect that’s a long-run, task-dependent number. But assuming he’s correct, that means humans have significant percentages of false rejects as well as false accepts. But how do you get your hands on this data?

One way would be to quarantine all inspected product and reinspect it. But with any respectable sample size your reinspectors will also have an error rate, so how do you know who is right?

I’m sure there are some statistical ways of handling this, but I’d like to suggest a more pragmatic approach: find out what your customers are telling you. Your Quality department almost certainly has data on returns and complaints: go through this to find out how much bad product human inspection is allowing through.

Calculate what that is as a percentage of total production and you have your False Accept rate. Subtract that from 20% and you have a False Reject rate.

Okay, that’s hardly scientific, but it will give you an indication of how much good product is being scrapped. (Yes, it’s probably over 10% because, in my experience, Inspectors tend to err on the side of caution.)

So how can you use this in your next project? Well hard numbers on the accuracy of your current performance will let you show how the vision system you install does a better job. (And if it doesn’t do better … well just make sure it does!) I suggest though, that rather than trying to design and run your own experiment out in the plant, ask Quality and HR to work on it for you. This will help get some buy-in and add credibility to the results. Sharing the numbers you estimated from customer complaints and Juran’s 80% rule might be enough to spur them into action without further nagging.

That’s what I mean by putting a stake in the ground: establish the current level of performance, so that you can, post-implementation, show an improvement.

Results of the touch screen survey

2011-08-23T22:42:00.000-04:00

A little over a week ago I asked, “Where do you stand regarding touch screen interfaces for machine vision?” Here’s what you said:

Love them and will use them everywhere 7%
Have used when the application warranted, and will do so again 60%
Tried one, didn’t like it 7%
Never tried using one for machine vision 26%

A couple of you also posted some interesting comments. One suggested that gesture control will rapidly become ubiquitous, rendering the mouse obsolete, while a second suggested buying a touchscreen monitor rather than a touchscreen computer.

Leaving aside the vision of programmers waving their arms around like some demented baseball player signaling his teammates, what are we to conclude from all this?

Well overall, it seems those who have tried a touchscreen found it useful, so if you’re in the 28% that have never given one a go, try it. However, there seems to be an implication in the responses that: a) they might work better for simple applications, and b) don’t splash out on a fancy touchscreen computer; just buy a monitor.

Waiting for your feedback!