Machine-vision focus shifts with application, speed and resolution remain key in AOI
-- Test & Measurement World, 1/8/2008 12:41:00 PM
Dalsa sales and marketing VP Philip Colet comments on hot topics in the machine-vision industry in an interview with T&MW chief editor Rick Nelson.
Q. What are the key issues facing the machine vision and inspection industry at this point?
A. When I look at machine vision, I look at the different components: lighting, optics, sensor technology, camera technology, data-acquisition technology, and image analysis and processing. Depending on type of application we are talking about, we will stress different components in different ways. If we talk about semiconductor applications, as line widths get smaller and smaller, we then in terms of illumination want to start talking about UV and deep UV, and that has implications for optical technology and sensor technology. And since the data rates are so high, there are also implications with respect to camera speed and data processing. If we talk about other applications—for example, traffic monitoring—the emphasis will be on different factors, like dynamic range. Speed of capture would not be as important as image quality under varying lighting conditions, which puts an emphasis on sensor technology.
Q. Where is the emphasis on PCB inspection applications—on frame rates and resolution, or on algorithms that can do more with less data?
A. In AOI applications, the throughput of a machine is a competitive advantage of the company making that machine, so the faster that machine works the better. The machine must of course analyze the images it acquires, but it all comes down to how quickly it can scan a PCB and how many PCBs it can do in an hour. So the AOI OEMs are asking for higher resolution, and they are asking for higher frame rates.
Keep in mind, there are a couple of different segments within electronic inspection: there’s bare PCB, and there’s assembled PCB. Start with bare—typically what’s used is a line-scan array. These applications are requiring higher resolutions because the line widths of the copper tracings are getting smaller and smaller as component pitches get a lot tighter. But in addition to the higher resolution, you want to be able to scan those PCBs a lot faster. So for bare PCBs we have a fairly easy equation—higher resolution and faster line rates is what’s driving those guys. These are monochrome applications—there’s no talk of color or color construction.
In the populated PCB world, OEMs are using high-resolution area-scan cameras—the 4-Mpixel node is very popular. And a speed of anywhere from 30 to 90 fps is also very popular right now. Where we see this going in a couple of years—and this is reflected in our product development effort—is to increase resolution up to 9 or 10 Mpixels and at the same time increase the frame rates. But, unlike for bare PCBs, we face a very complicated question, because we have a multivariable equation describing a relationship between resolution, the size of the sensor (which then determines the size of the pixel), the lighting, the frame rate, and the optics. You have to find the optimum combination of all of those things—along with the right price.
Q. Do populated PCB AOI systems employ color?
A. A machine typically generates two images: a monochrome image and color image. The monochrome image is used to inspect solder joints, looking for bridges between leads and also inspecting the profile of each solder connection. Then the color image is used to look at components themselves to see if there are things like tomb-stoning or components put in the wrong way.
Q. Dalsa recently introduced the Falcon 1.4M100 area-scan camera. What are some of its features?
A. Its resolution of 1400 by 1k pixels is good for a variety of machine-vision applications. Also, the size of the sensor ties in nicely to the size of available lenses, so you can use a relatively inexpensive lens, and you can get relatively close to your object. But you can get other sensors like that. The unique feature about this camera is the speed—the ability to go to 100 fps at that resolution. And that speed of image acquisition can be useful in, for example, the populated PCB AOI we were just discussing. It can also find use in the high-speed chip shooters that are populating PCBs, and it can also be used in package inspection applications, where you need to inspect different objects coming down a production line and be able to capture images at production rates.
Q. Is it a CMOS or a CCD camera?
A. It’s a CMOS sensor that provides very good quality images with quite a good signal to noise ratio as well as quite good CCD-quality light-gathering capability. The whole debate about CMOS vs. CCD is to us almost irrelevant. To us, the image quality we can get with CMOS sensor is just as good as what we can derive with many CCD imagers on the market today.
Q. So CMOS will completely replace the CCD?
A. CCD won’t go away—there will continue to be niche applications for CCD—for example, for applications requiring backside thinning. Backside thinning is a lot easier to do with CCDs than with CMOS. A second example is the TDI [time delay and integration] sensor. Dalsa makes a lot of TDI sensors, which are essentially line-scan sensors where you have this bucket brigade in which you are dumping charge. With CMOS that just doesn’t happen, because you don’t have a charge anymore, you have a voltage, and you can’t really accumulate voltage the way you can accumulate charge. So TDI doesn’t work with CMOS. So our feeling is, CCD will stay in these niche applications, but CMOS will take over in terms of image quality. We are already starting to see more commercial-grade digital cameras coming out with CMOS sensors, because the benefits really are there with CMOS technology: mass production, single-voltage power supplies, lower power consumption, and the ability to window easily.
Q. What’s happening with image processing—is it still taking place on a dedicated image-processing board, or is it migrating to the PC?
A. You have to look at progress in the industry in terms of image processing. If you go back 15 years you had relatively lame computers, and all of the processing had to occur on special image-processing hardware. But then two things happened. Number 1, the PCI bus became popular, and it allowed you to dump raw data down to PC. And number 2, we saw a corresponding rise in the processing power of the PC. It was those two things happening at the same time that enabled a transfer of processing responsibility from an onboard architecture down to a standard PC architecture.
Q. So the embedded architecture is dead?
A. No way. Because people are always pushing the envelope—with the kinds of sensors that are coming out right now, the data rates they are running at and the complex algorithms the people want to execute can still overwhelm even a quad-core PC, so you still have to do some processing on the board. Where do we find those applications? They tend to be in anywhere you are generating a lot of data—such as semiconductor wafer-inspection applications—or they tend to be in real time applications, where I do not have the ability to accumulate data and then process it. Those tend to be medical applications where there is a tremendous amount of data. You may have a 4k by 4k sensor that’s running at 30 to 60 fps with 12-bit dynamic range. As for processing, you may want to normalize the image, you may have to do some noise reduction on that image, which may require some kind of adaptive algorithm, and you may have to account for pin-cushioning and other artifacts. In addition, you may have to rotate the image at an arbitrary angle. When you add up all these different steps, you’ll find they will overwhelm any PC that’s on the market today, so you have to do that with an embedded type of application, doing the processing in FPGAs, for example. And it has to be real time. Because for surgeons or doctors who are performing a procedure, the last thing they want to have is a two-second lag between them manipulating something and then seeing the results on a monitor.
Q. Is programming difficult with an embedded FPGA architecture?
A. There are a couple of approaches to that problem. Any time you look at FPGA programming, it’s a fairly onerous task if you have to get down to the level of the HDL coding. What we do is package FPGA-based image-processing algorithms for the OEMs so that they don’t have to worry about what’s under the cover. They can just program those items from host in software and rely on the fact that we will port those algorithms from one generation platform to the next so they can preserve their intellectual property. If OEMs need to get down to the level of custom coding, then we open up the hood for them so they can do that. We just try to make sure that they have people on staff who can do that kind of work, or we undertake to do that for them as a service.
Q. What will be the hot topics in machine vision in 2008?
A. The debate will continue between CCD and CMOS, and I expect CMOS to continue to progress in terms of image quality and light sensitivity. In terms of camera technology, I expect the issues to have more to do with the interconnect as opposed to what’s in the camera itself. The hot subjects there would be where we are in terms of 10 GigE Vision and where we are in terms of Camera Link and its limitations. We have the mini Camera Link connector now and power over Camera Link, which are real benefits to OEMs, but what comes next? Camera Link will bring you up to about 680 Mbytes per second maximum, but we are talking about sensors that will be going at a gigabyte per second or even 2 or 3 gigabytes per second. What interconnect technology are we going to be using for them? We will be looking at PCI Express to see what it can do for us. We will be looking at FPGA technology and how it might enable preprocessing that extracts relevant information from the camera. And we will be looking at how we ensure the integrity of the data that’s being generated by the camera sensor and transmitted to the host.
In terms of software technology, there will be continued work on pattern recognition and on color. Today, color is complicated because there are a lot of ways to look at color, and color tools now are somewhat antiquated in how they force people to think about color. New color tools should simplify the life of the OEMs so they are not forced into thinking in one domain or another.
Talkback
Related Content
Related Content
Sponsored Links
TMW Resource Center
Browse Resources by Type:
- The Ultimate Infrared Handbook for R&D Professionals
FLIR Systems | Reference Guide
VIEW NOW
Cincinnati Sub-Zero | White Paper
VIEW NOW
Cincinnati Sub-Zero | Reference Guide
VIEW NOW
View More
