FlexInspector inspection methods (1)

I would like to introduce FlexInspector's inspection method again.
Basically, FlexInspector registers images of good items and detects the areas that are different from the good items. Foreign objects, dimensional differences, and scratches are all detected as "areas that differ from the good", and as a result, only "the same as the good" is judged as OK.
The problem here, however, is the "variation of good items.
Since it is impossible for a single good image to be exactly the same as another good image, a "difference" algorithm that assumes this will not work well. To cope with this, there are two possible methods: one is to obtain the average from multiple images and use it as the master, and the other is to compare it with multiple good images. The former, however, has the problem of how to define the range of variation. The latter method is time-consuming because multiple images are compared. The question is whether the "variation" can be absorbed by a few good images at most.
FlexInspector statistically processes multiple good images to obtain the mean luminance and standard deviation σ for each pixel, and then determines the upper and lower luminance limits for each pixel. Then, during the inspection process, it checks whether the image is within the upper and lower limits. Technically speaking, it is not so much "matching" as it is "the ultimate form of binarization" in that each pixel has its own optimized binarization level.
It is like HALCON's VariationModel! That's what you say, but... after 6 years of figuring out this method, it may not be the answer if we keep it that way...