Insulator defect detection based on improved Yolov5s

Abstract

The stable operation of a power supply system is inseparable from the work of detecting defects in transmission lines. However, the insulator defect detection model based on deep learning is widely used in wire inspection work. Therefore, this paper proposes an improved YOLOv5s insulator defect detection model in order to solve the problems of insufficient training data and low recognition accuracy of the target detection model in the real-time detection of small target insulator defects. To expand and enhance the training data, experiments were conducted using the addition of noise and random black blocks. The spatial and channel weight coefficients were obtained by adding an attention mechanism (Convolutional Block Attention Module, CBAM), and the dimensions of the input feature maps were transformed to enhance the model’s ability to extract and fuse small target defect features. Experiments show that with Faster RCNN, YOLOv3, SSD and YOLOv4 comparison experiments verified that the algorithm achieves 97.38% detection accuracy for insulators and 93.32% detection accuracy for small target insulator defects with a fast detection speed, which is a better solution to the problem of detecting insulator defects with too small a proportion in the image.