Surface defect detection method for high-voltage cable joints based on improved YOLOv5s

Aiming at the problems of inaccurate and time-consuming manual detection of surface defects on high-voltage cable joints, a method for detecting surface defects on cable joints based on improved YOLOv5s is proposed. Firstly, (Space to Depth) SPD convolution and (Efficient Channel Attention) ECA attention mechanism are used in the feature extraction network to prevent the loss of information in tiny defects due to pooling operations. Secondly, a new multi-scale fusion feature pyramid structure is designed to better retain and fuse important information at each scale. Then the decoupling operation is performed on the detection head to avoid the conflict between the classification task and the localization task. Finally, the (Normalize Wasserstein Distance) DWN metric is used to design the localization loss to improve the ability of tiny defect localization. The experimental results show that the proposed method achieves 93.86% mAP in the cable joint surface defect dataset. Compared with the original algorithm, it improves 2.61% mAP, and at the same time reduces the number of parameters by 27.46% and the flops by 49.43%.