Cross-attention fusion and edge-guided fully supervised contrastive learning network for rail surface defect detection

Abstract

In recent years, there has been significant research focus on efficiently and accurately detecting defects on rail surfaces using computer vision. Utilizing depth information from the rail surface has emerged as an effective approach for detecting visually insignificant types of defects that are unique in nature. However, previous methods have typically overlooked the long-distance dependency between the two modalities when fusing them using conventional convolutional network methods. Additionally, these methods have often relied on traditional cross-entropy loss for edge supervision without considering the intra and inter-pixel relationships associated with edge features. To address these limitations, we propose a novel approach called CECLNet (cross-attention fusion and edge-guided fully supervised contrastive learning network) for rail surface defect detection (RSDD). The proposed CECLNet incorporates a module for inter-modal cross-attention fusion, which effectively explores the complementary information by considering the long-range relationship. Furthermore, we introduce a progressive aggregation-based multiscale feature interactions decoder to promote sufficient information interaction between multiscale features, thus facilitating the generation of final predictions. Finally, we propose a pixel-level fully supervised contrastive learning approach to enhance the efficiency of utilizing edge-assisted information. Extensive experiments conducted on the industrial NEU RGB-D RSDDS-AUG dataset demonstrate the superiority of our proposed CECLNet over 17 state-of-the-art methods.