Hierarchical Cross-Modal multianchor distillation for rail surface defect detection

For red-green-blue with depth (RGB-D) rail surface defect detection (RSDD), most models focus on improving accuracy, while ignoring computational costs. To address this gap, we introduce a hierarchical cross-modal multi-anchor distillation network (HCMD) for RSDD. Moreover, we propose an innovative defect detection approach, comprising three key stages: merge, split, recombine. During feature extraction, multi-layer features are divided into high-and low-level guidance features using a cross-modal cross-stage fusion module. Subsequently, the learned semantic feature is used to guide the low-level feature fusion using a multiscale graph convolution. Finally, to reduce the computational effort, hierarchical cross-modal multi-anchor distillation loss is used to transfer the teacher's knowledge to the student, thus achieving model compression. Extensive experiments using the NEU RSDDS-AUG benchmark RGB-D dataset demonstrate our HCMD's superiority in terms of reduced computational complexity and high accuracy.