Interactive Fusion Distillation Network Embedded With Poisson Images for Wheel Tread Defect Detection

As an essential component of the train system, the wheelset is prone to defects on its tread surface due to the influence of high-frequency impacts during operation. However, due to the complex and changeable wheel-rail contact conditions, tread defects face problems such as complex backgrounds, diverse scales, and small defects. An interactive fusion distillation network embedded with Poisson-coded images (RP-IFDN) is proposed to address the above issues for tread defect detection. Specifically, a dual backbone network (DStarNet) embedded with Poisson coded images is designed to extract and fuse the texture features of red, green, blue (RGB) images and the edge features of Poisson (POS) images, thereby alleviating the problem of complex background; An adaptive interactive squeeze-excitation (AISE) module is designed to extract the interactive information between RGB features and POS features and enhance the focus on essential areas, thereby improving the efficiency of feature fusion and suppressing background information; A global-context feature pyramid (GC-FPN) is proposed to capture global-context information when fusing cross-layer features, thereby enhancing the ability to detect small target defects and extract multi-scale features; A straightforward knowledge distillation strategy is proposed to transfer the knowledge from a dual-backbone detector to a single-backbone detector. This approach addresses the challenges of excessive parameters and computational complexity associated with dual-backbone architectures. The effectiveness of the proposed method has been demonstrated through a series of comparative experiments, ablation experiments, visualization analysis, and generalization performance experiments.