Weakly Supervised Surface Defect Detection Based on Attention Mechanism

Abstract

<p indent=0mm>Nowadays, many defect detection algorithms based on deep learning are trained using a supervised learning strategy, which largely depends on the number of required samples and the quality of annotations. A weakly supervised attention network for surface defect detection is proposed, named SDD-Net, which can simultaneously predict both the location and probability of defects only by using image-level labels. Firstly, the feature fusion module is used to extract multi-scale edges features from the output of the multi-scale receptive field. Then, the deep semantic information of features is mined by multilevel auto-coder. Meanwhile, the trilinear global context attention module is used to further refine the spatial location information of shallow features. Finally, the SDD-Net is used to integrate the shallow edge features and deep semantic features to obtain the final fine defect features. The results of evaluation on KolektorSDD dataset demonstrate that the SDD-Net based on the PyTorch framework has better detection performance than other detection methods such as U-Net. It can retain more detailed texture information and effectively expand the feature difference between small defects and complex background. The experimental results show that the proposed model is more accurate than the other models in the complex scene, the precision, the <italic>F</italic>₁-score and the classification accuracy are significantly improved.