Reliable and Robust Weakly Supervised Attention Networks for Surface Defect Detection

Abstract

Automated surface-anomaly detection of industrial products using deep learning is a critical task to the digitalization and intelligence of the manufacturing industry. In recent years, with the rapid development of artificial intelligence techniques, deep learning has been successfully applied to this task. However, the current defect detection methods based on deep learning usually adopt strong supervised learning strategy such as object bounding box or pixel-level labels to predict the location of defects, which leads to the performance of the algorithm depends on the number of data provided and the quality of the annotations. Therefore, how to significantly reduce the cost of data annotation without reducing the performance of the model, that is a challenging and urgent task. As such, this paper proposes a weakly supervised attention network designed for surface defect detection. It can simultaneously predict both the location and probability of defects only by using image-level labels. Experimental results also demonstrate that the proposed method is able to learn on a small number of surface defect data, and can accurately realize automatic evaluation of defect detection, showing great potential for industrial application.