Anomaly detection with convolutional neural networks for industrial surface inspection

Abstract Over the recent years Convolutional Neural Networks (CNN) have become the primary choice for many image-processing problems. Regarding industrial applications, they are hence especially interesting for automated optical quality inspection. However, with well-optimized processes is it often not possible to obtain a sufficiently large set of defective samples for CNN-based classification and the training objective shifts from defect classification to anomaly detection. Here we approach this problem with deep metric learning using triplet networks. Our evaluation shows promising results that even translate to novel surface/defect classes, which were not part of the training data.