Appearance Inspection Algorithm Contest

Did you know that there is an Appearance Inspection Algorithm Contest?
Please take a look here.
The contest is sponsored by the Japan Society for Precision Engineering and has been going on since 2001.
The theme of this year's contest is "Detection of Macro Defects on Semiconductor Patterns.

As a self-proclaimed "visual inspector" I was also interested and ordered some images.

What I thought at the

If you have a good image, it is not difficult. It is a piece of cake if you can position and compare them.
However, the difficulty level goes up dramatically when you are shown only an image and asked to find something wrong with it, as in this contest. Even if you work hard, it is difficult to achieve perfection.

Certainly, it is meaningful to consider algorithms for visual inspection in terms of research on the "human thought process. However, in most cases, there are "good products" in the visual inspections required at manufacturing sites. Also, 100% is not a contest in terms of defect detection capability. 99 points is also not good enough.

The preconception that visual inspection "detects defects" is too strong. Because of this preconception...

VISUAL INSPECTION IS "TO DETECT GOOD PRODUCTS. IT IS SUFFICIENT TO MECHANICALLY DETECT ONLY THOSE ITEMS THAT ARE IDENTICAL TO GOOD PRODUCTS. IF THERE IS SOMETHING DIFFERENT, IT SHOULD BE LEFT AS NG FOR NOW, AND A HUMAN BEING CAN MAKE A JUDGMENT LATER. THE STORY BECOMES DIFFICULT BECAUSE WE ARE TRYING TO MAKE A DIFFICULT JUDGMENT BY A MACHINE. IF THE MACHINE DOES A POOR JOB AND OUTFLOWS DEFECTIVE PRODUCTS, IT WILL BE A LOSS FOR ALL CONCERNED.

As a research theme, it would be much easier and more meaningful to master "what a good product is" than to pursue unidentifiable "defects.