LINX DAYS 2021 SUPPLEMENT

AT "LINX DAYS 2021," WHICH WILL BE HELD ONLINE ON NOVEMBER 25 AND 26, OUR CASE STUDY WILL BE PRESENTED UNDER THE TITLE "BEYOND RULE-BASED! OVER-DETECTION CONTROL WITH ANOMALY DETECTION" WILL BE PRESENTED AT LINX DAYS 2021.

Frankly, there have been quite a few changes by Mr. Links...

Apparently you want to use the term "rule-based" as a synonym for "deep learning".

IN OUR COMPANY, "RULE-BASED" IS AN APPROACH THAT USES IMAGE PROCESSING TO DEFINE DEFECTS. WE BELIEVE THAT ITS SYNONYM IS "GOOD COMPARISON," AN APPROACH THAT CONSIDERS A DEFECT OK AS LONG AS IT IS THE SAME AS A GOOD DEFECT.

The purpose of this issue is that "good comparison" using anomaly detection by deep learning, as opposed to "good comparison" using the conventional method (VariationModel), can process good variation more flexibly, and thus over-detection can be greatly suppressed.

RULE-BASED" IS AN APPROACH THAT DETECTS ONLY DEFINED DEFECTS, SO OVER-DETECTION IS UNLIKELY TO OCCUR, WHILE "GOOD PRODUCT COMPARISON" IS AN APPROACH THAT "DOES NOT JUDGE OK UNLESS THE DEFECT IS THE SAME AS A GOOD PRODUCT," I.E., IT IS DIFFICULT TO TOLERATE VARIATIONS IN GOOD PRODUCTS, SO OVER-DETECTION IS LIKELY TO OCCUR. IT IS A BIG ADVANTAGE TO BE ABLE TO SUPPRESS THIS.

WE HAVE RECEIVED SEVERAL THOUSAND IMAGES OF GOOD PRODUCTS FROM USERS OF OUR EXISTING "GOOD PRODUCT COMPARISON" SYSTEM FOR EVALUATION. IN ONE CASE, WE LEARNED MORE THAN 7,000 IMAGES OF GOOD PRODUCTS, RESULTING IN A 99.3% GOOD PRODUCT RATE AND A 100% DETECTION RATE OF NG SAMPLES THAT SHOULD BE DETECTED. ON THE OTHER HAND, IN ANOTHER CASE, WE FOUND A PHENOMENON IN WHICH FINE DEFECTS IN AREAS WITH LARGE VARIATION COULD NOT BE DETECTED.

If you have many good images, we can easily evaluate them, so please contact us.