A. The primary challenges for vision suppliers are robustness, speed, and ease of machine set-up. In terms of robustness, getting the right answer is job one. It’s extraordinarily challenging to get right answers because perfectly good product varies enormously in appearance. There are manufacturing process variations, surface reflectance, not to mention that feature sizes keep shrinking. Because there’s so much variability, it’s extremely hard to tell what is and what is not acceptable. Vision systems need to be robust enough to locate patterns, despite the many pattern variations and must distinguish between the acceptable and non-acceptable.

The second major challenge is speed, or how fast you need to get the right answer in semiconductor manufacturing. Getting the right answer in the tens of milliseconds that are available to do the job is critical, and those times are becoming shorter as the industry matures.

Finally, reducing the complexity of machine set-up is a huge challenge. This set-up is typically a time-consuming process that may involve very costly downtime. Vision systems today are easier to install, and reduce the time spent on machine set-up.

Q. The lines and spaces on today’s integrated circuit packages are becoming denser. Is there a finite level beyond which a vision system cannot see?

A. When dealing with the practical limitations of today’s shrinking geometries, there are optical considerations, sensor considerations and algorithmic considerations. Optics are the easy part, since they are nowhere near the limit of their ability to resolve alignment and inspection problems.

With regard to sensor considerations, we may start seeing higher resolution sensors coming on the market, but overall resolution has not appreciably changed in the last 40 years. The most significant changes have taken place within machine vision algorithm development and the ability to extract information from images. In the early days of vision, we could extract to a single pixel. Then, in the late 1980s we improved this to extract to _ pixel. The big advance in the last few years has been the ability to extract image data to 1/40^th pixel and maintain that under rotation and size changes.

As machine vision algorithms continue to improve, and more and more information is able to be extracted from images, the limits will be a function of the information actually present in the manufactured objects themselves and how consistently patterns are produced. We recommend that semiconductor manufacturers work directly with vision suppliers so they will understand the needs of visual inspection and design products--not just for manufacturing, but for automated visual inspection and alignment, as well.

Q. As vision systems become smarter, do you anticipate that their cost will increase, decrease or stay the same?

A. History has shown that the cost of vision is decreasing, especially when you look at the low-cost vision sensor technology that is now available. When the industry started, everything was roughly around the same price. Perhaps more important than whether prices will go up or down is the fact that the semiconductor industry is seeing, and will continue to see, a much richer variety of price/performance points for vision technology.

Q. Machine vision has for many years proven its worth in aligning, inspecting, and identifying semiconductor components throughout the manufacturing process. What is the vision industry doing to help classify defects?

A. We recognize defect classification as one of the next frontiers for vision, aimed at providing substantive improvements in manufacturing efficiency. What has limited vision in this area is the number of acceptable parts and the almost infinite range of what could be classified as "defective." This presents a very difficult problem, even for human experts who often have trouble getting beyond 60-80% accuracy in distinguishing between defects and acceptable variations. We see promise, however, in new algorithms that aid in defect classification, which help manufacturers understand the types of defects being found.

Q. Have the applications for machine vision in semiconductor manufacturing changed significantly over the years?

A. One of the interesting things about machine vision in semiconductor applications is that we are still working to improve things that we were working on in 1981. For example, the very first product we introduced at Cognex was a wafer reader, and today we are still working on improving wafer reading. We don’t simply take on a job, solve the problem and move on. Rather, there is a continual development and improvement cycle that occurs as processes and requirements in the industry change.

Q. How much of vision inspection is hardware and how much is software? Is the proportion changing?

A. A machine vision system that performs inspection (or other tasks), typically consists of an industrial camera, processing hardware, and vision software. Historically, the software consisted of vision tools running on dedicated hardware. However, this has changed as the computing power of PCs has increased. Since today's PCs offer an excellent price/performance ratio, many vision applications now run on PCs. The success of the PC has allowed us to focus more on providing superior vision tools, with much less time spent on the hardware platform necessary to deliver these tools.

While cameras and hardware play an important role in vision inspection, the key to successful practical applications is the software, which acts as the brains of a vision inspection system. It is what guides the equipment and inspects, gauges or identifies the ICs during the semiconductor assembly process. Although humans do not have the best eyesight, are not the fastest or strongest species, we do possess superior software (the brain) that is the key to our intelligence and advanced status. As with humans, software is also the key to an inspection system's intelligence.