Anomaly score is very low for unseen defective images

[InterKnight]

Exciting for such an excellent project. I tested it with my own dataset and found that the model performs exceptionally well at detecting labeled defective images it has been trained on, and it does not produce false positives for normal images it has never seen before. However, its detection capability is weak for unseen defective images. Specifically: although the abnormal regions are correctly highlighted in the heatmap, the anomaly score is very low. Could you please offer some suggestions for improvement?

[blaz-r]

Hi, great to see you like our project.
I'm not too sure about this issue, so could you provide more details? What config do you use? Ideally, for supervised, I'd say the one for Sensum is used. How long did you train and what's the data like (number of anomalous and normal, etc).