TLEAR‐Net: A Network for Defect Detection in Train Wheelset Treads Based on Transfer Learning and Edge Adaptive Reinforcement Attention

ABSTRACT As a critical load‐bearing and running component of railway systems, the wheelset's operational safety fundamentally depends on precise detection and localisation of tread defects. Current deep learning‐based detection methods face significant challenges in extracting discriminative edge features under small‐sample conditions, leading to suboptimal defect localisation accuracy. To address these limitations, this study proposes TLEAR‐Net, a novel defect detection framework integrating transfer learning with an edge‐adaptive reinforcement attention mechanism. The methodology employs RetinaNet as the baseline architecture, enhanced through multi‐stage domain adaptation using COCO 2017 pretraining and parameter‐shared ResNet‐50 backbone optimisation to bridge cross‐domain feature discrepancies. An innovative edge‐adaptive reinforcement (EAR) attention module is developed to selectively amplify defect boundary features through learnable gradient operators and hybrid spatial‐channel attention mechanisms. Comprehensive evaluations on a proprietary data set annotated defect samples demonstrate the framework's superior performance, achieving state‐of‐the‐art detection accuracy (89.22% mAP) while maintaining real‐time processing capability (42.45 FPS).