IS AI (DEEP LEARNING) NECESSARY FOR APPEARANCE INSPECTION?

It depends on the object.

IF ALL THE ITEMS TO BE INSPECTED ARE MADE IN THE SAME WAY, IT IS EASY AND RELIABLE TO INSPECT THEM TO SEE IF THEY ARE THE SAME AS THE GOOD ITEMS. THERE IS NO EASIER AND MORE RELIABLE WAY.

The problem is that good products also have variations. Since this variation causes variation in the images of good products as well, simply superimposing them over each other is not the way to deal with this problem. There are statistical methods and deep learning as a way to deal with this variation.

The statistical method is an inspection method that we have employed for nearly 20 years and is still very widely used today. The variation of a good image is defined as the variation of luminance values at the same coordinates, and the rule of "average ±3σ" is applied to each pixel to generate a lower limit image and an upper limit image as a good image. The system has a very high detection capability for objects with a small variation σ and for areas with a small variation, and as the variation σ increases, the range of good products increases, resulting in a loose inspection. However, the tolerance of variation is not large, and the inspection is suitable for high-precision objects, and the percentage of good products tends to be low. Instead, defective products will not be mixed in with those judged as OK.

DEEP LEARNING CAN BE THOUGHT OF AS A WAY TO TRY TO DEFINE THIS VARIATION IN GOOD PRODUCTS BY HAVING AI LEARN IT. IT HAS THE POTENTIAL TO INSPECT GOOD PRODUCTS EVEN IF THE IMAGES OF GOOD PRODUCTS ARE NOT IDENTICAL. IT HAS THE POTENTIAL TO BE APPLIED TO ALL KINDS OF VISUAL INSPECTIONS, INCLUDING THOSE WITH SMALL VARIATIONS, BUT THERE ARE SEVERAL PROBLEMS WITH STATISTICAL METHODS.

• CERTAINTY. THE GREATER THE TOLERANCE FOR VARIATION, THE HIGHER THE PERCENTAGE OF GOOD PRODUCTS. IN OTHER WORDS, THE RANGE OF GOOD PRODUCTS IS INCREASED, AND THE PROBABILITY THAT AN OK-RATED PRODUCT IS A GOOD PRODUCT IS DECREASED.
• Immediacy of learning. In the case of statistical processing, the process of registering additional good products and updating the upper and lower limit images is instantaneous. Deep learning, on the other hand, requires several hours of relearning.
• HARDWARE LIMITATIONS. DEEP LEARNING USES GPUS FOR LEARNING. IN ORDER TO TRAIN IMAGES FROM 2M AND 5M CAMERAS, WHICH ARE COMMONLY USED IN THE MARKET TODAY, A GPU WITH MORE THAN 20GB OF MEMORY IS REQUIRED, WHICH IS A SIGNIFICANT COST INCREASE. THIS GPU MEMORY CAPACITY ISSUE MAKES IT DIFFICULT TO BUILD SYSTEMS WITH ULTRA-HIGH PIXEL CAMERAS OR MULTIPLE CAMERAS.
• Software Issues. This is a rapidly changing technology, and the environment may have changed dramatically in a year's time. If you use commercial products, you can expect manufacturer support, but on the other hand, you will also have to pay licensing fees.
• Cost issue. As a rough guide, 1.5 million for statistical methods and 3 million for deep learning; a cost increase of about 1.5 million yen is unavoidable. Considering a budget of 10 million yen for automation, this is a fatal cost increase.

Considering these problems, "deep learning for everything" is not a good idea. For objects with large variations that cannot be handled by statistical methods, deep learning may be the best approach at this point.

One last thing. If we are going to invest in deep learning for inspections to deal with variation in good products, wouldn't it be more essential to invest in process improvement to reduce variation in good products? If the variability of good products is reduced, the frequency of defective products should also be greatly reduced.