How to improve the ability to detect missed abnormal images?

[surlovely]

In patchcore, for practical applications, if a positive sample is overchecked, I can add the overtested sample to the training set. But when it comes to missing an anomalous sample, there seems to be nothing I can do. Do you have any ideas?

[alexriedel1]

You could add the anomalous sample to the abnormal validation data. So the threshold might be better adjusted during validation.

[abc-125]

You can also try another model; it might work better than PatchCore (although it depends on your specific case, and the chances are not high, PatchCore is the best freely available model for now). I would suggest Reverse Distillation (available in Anomalib) or GLASS (not available in Anomalib yet).