Research on Tow Surface Defects Detection Technology Based on YOLOv5

In the process of tow manufacturing, due to the high consistency of the color and material of the tows, the traditional manual detection method often has the problem of missing detection and false detection. Aiming at defects in tow images, an improved YOLOv5 tow defect detection algorithm is proposed based on YOLOv5 algorithm. Firstly, ODConv dynamic convolution is used in the backbone network to refine the local feature mapping to achieve full extraction of local fine-grained features of the strands, and the convolution layers and BN layers are fused to reduce the number of network layers and improve the problem of reduced computing efficiency caused by the addition of dynamic convolution. Secondly, the Biformer attention mechanism is introduced into the Head structure to retain the fine-grained details of the detected object as much as possible. Next, in order to further improve the detection effect of small targets, a new bounding box similarity measure (Normalized Gaussian Wasserstein Distance, NWD) is introduced into the loss function to combine with the original IOU_Loss and improve the convergence speed. Finally, experiments are carried out on tow defect datasets to verify the performance of the improved model. The conclusion shows that although the number of parameters and calculation time of the improved model are increased, the detection effect of small target defects of the tow is improved, and the detection requirements of tow defects in the production process are better met, which provides a new method and idea for the research of tow detection in the field of deep learning.