Explaining Defect Detection with Saliency Maps

The rising quality and throughput demands of the manufacturing domain require flexible, accurate explainable computer-vision solutions for defect detection. Deep Neural Networks (DNNs) reach state-of-the-art performance on various tasks but wide-spread application in industrial is blocked by lacking explainability DNN decisions. A promising, human-readable solution given saliency maps, heatmaps highlighting image areas that influence classifier’s decision. This work evaluates a selection methods area assurance. To this end we propose distance pointing game, new metric to quantify meaningfulness maps We provide steps prepare publicly available dataset defective steel plates proposed metric. Additionally, computational complexity investigated determine which could be integrated edge devices. Our results show DeepLift, GradCAM GradCAM++ outperform alternatives while cost feasible real time applications even indicates respective used as an additional, autonomous post-classification step explain decisions taken intelligent assurance systems.