Improved swin transformer-based defect detection method for transmission line patrol inspection images

Correctly locating transmission line defects and taking timely remedial measures are essential to ensure power systems' safety. Convolutional neural networks (CNNs) are commonly used in defect detection in transmission line inspection images, but the local nature of the convolution operation limits the detector's performance. Transformers have become more and more prominent in the field of computer vision because of their global computing function. This paper proposes a transmission line image defect detection method that combines CNN and Transformer comprehensively. In particular, an enhanced local perception unit is designed to reduce false and missed detections of small and occluded objects. The problem of the high computation and complexity of the Multi-Head Self-Attention module is solved via a lightweight self-attention method. In addition, an adaptive multi-scale fusion module is designed to extract more effective fusion features and improve the model's robustness. The numerical realization of the proposed method versus Faster Region-based Convolutional Neural Network (Faster R-CNN), Cascade R-CNN, DEtection TRansformer (DETR)-R50, You Only Look One-level Feature (YOLOF), You Only Look One X-Large (YOLOX-L) and Swin Transformer (Swin-T) proved its superiority in the average accuracy of transmission line image defect detection.