SSDC-Net: An Effective Classification Method of Steel Surface Defects Based on Salient Local Features

To effectively classify steel surface defects, it is crucial to improve the feature representation. The features extracted by the current steel surface defect classification model inadequately represent different defect categories, particularly when samples from different categories exhibit high similarity, resulting in degraded model performance. In this paper, we propose the Steel Surface Defect Classification Network (SSDC-Net) to optimize the steel surface defect classification task. This network includes the Channel Information Complementary Matrix (CICM) module and the Multi-Scale Contrastive Loss (MCL) method, which models the channel information between similar samples to capture and distinguish subtle visual differences induced by distinct channel information. Specifically, the CICM module utilizes interactive information between channels of the feature map to enhance local feature representation. To enhance global feature representation, we further introduce a multi-scale contrastive loss, which not only maximizes differences between features but also reduces the gap between different views of the same sample. Extensive experiments conducted on the BG-DET dataset, FSC-20 dataset, and FSD dataset demonstrate the effectiveness of each proposed module. Compared with other methods, our approach has achieved state-of-the-art performance. Additionally, our method is plug-and-play and can be seamlessly integrated into other backbone networks.