Generative Models and Quality Constraints for Anomaly Detection: Application to Industrial and Medical Images

Anomaly detection in critical applications such as industrial or medical domains struggles to reach the lowest false positive rate in the zero false negative setting. Our work aims to reconstruct automotive PCBA or other industrial/medical images, through a Vector Quantized GAN (benchmarked with ViT and LDM generative models), so that if an anomaly is present, the reconstructed image shows the normal version of the product. After a collection of metrics on the residual image and other statistics, a neural network classify the image as normal or abnormal, through a high true positive rate optimization formulated by a partial AUC approximated by a Wilcoxon-Mann-Whitney statistics loss function.