YOLOv8-ECA-SPD-Lite for Cantilever Bolt Defect Detection in Railway Catenary Systems

To address the challenges of highly variable operating conditions in electrified railway catenary systems, the small target size of cantilever bolt defects, and the large parameter sizes of existing catenary inspection models, this paper proposes an improved lightweight algorithm named You Only Look Once version 8 Enhanced by Efficient Channel Attention, Space-to-Depth Convolution, and Lightweight Optimization via Pruning and Distillation (YOLOv8-ECA-SPD-Lite). The core innovations lie in: 1) Data augmentation on the original images to improve generalization capability; 2) Integration of the Efficient Channel Attention (ECA) module to enhance feature responses in bolt regions for small target detection; 3) The Space-to-Depth convolution (SPD-Conv) module was first applied to cantilever bolt defect detection to minimize feature loss for small targets; and 4) Replacement of the detection head with a decoupled head to enhance defect perception capabilities. Furthermore, applying model pruning and distillation techniques significantly improves model lightweighting, facilitating deployment on embedded onboard systems. Ablation studies validated the effectiveness of each proposed module. Comparative experiments demonstrated that the improved YOLOv8 model outperforms Faster R-CNN, SSD, YOLOv6s, YOLOX-s, YOLOv9s, and the original YOLOv8s across multiple metrics. Specifically, the proposed model achieved a higher mean Average Precision (mAP) than all comparison models. In terms of lightweight design, the improved YOLOv8s model achieves the lowest computational complexity among all compared methods, with only 17.2 GFLOPs. It demonstrates exceptional suitability for mobile deployment. This research holds significant theoretical and practical value, facilitating the transition of railway operation and maintenance to artificial intelligence, enhancing the efficiency of catenary inspection, and meeting the demands of high-density railway operations.