An Object Detection Model for Power Lines with Occlusions Combining CNN and Transformer

The application of computer vision in power line inspection has made some progress. However, the complexity of power line inspection images also brings some difficulties and challenges to this field. Aiming at the problems of separated features in power line occluded object detection, which may lead to missed detection and lower detection accuracy, this paper proposes a new power line detection model based on You Only Look Once X-s (YOLO X-s). Firstly, based on the Vision Transformer (ViT), a novel self-attention Transformer structure (SATS) is designed and integrated into the YOLO X-s structure, which enables the model to establish correlations with segmented features and capture global features, and enhance the detection performance of the occluded objects. Secondly, a Feature Connected-Feature Pyramid Network (FC-FPN) is proposed to replace the Feature Pyramid Network (FPN) and Path Aggregation Network (PAN), which can implement repeated interaction and fusion of both the global features and the local features. Thirdly, a Selective Kernel Network (SKNet) is selected to adaptively select two kinds of features with the same dimensions in the channel dimension through channel attention weights and finally output to the head structure. The results show that the average detection accuracy of the detection model proposed in this paper is 6.88% higher than the YOLO X-s model, reaching 82.00%, and the detection speed reaches 37.2 frames per second (FPS). The code is available at https://github.com/fancy7/ODNet.