WDD-YOLO:Detection of Small Weld Seam Defects Using Improved YOLOv5

As the weakest point of the pipeline, the weld seam is prone to various internal defects and has great safety risks, so it is necessary to conduct a weld seam inspection. In addition, accurately detecting small-sized defects is still a challenging task. This paper proposed a yolov5 network ramework for detecting small-sized defects using a mixed model that enjoys the benefit of both self-Attention and Convolution (ACmix). First, Asine function transformation was applied to obtain clearer weld images, and ACMix was added before the Spatial Pyramid Pooling (SPP) module in the backbone network to enhance feature extraction capabilities. SIOULoss was introduced to replace the original GIOULoss bounding box loss function, improving detection accuracy and training speed. Defect detection was conducted in the experiments using a custom-built dataset and the GDxray dataset, The experimental results show that, compared with similar models, the proposed network framework improved the precision of weld defect detection from 83.6% to 85.2%, increased the recall rate from 75.4% to 77.6%, and raised the Map@0.5 from 84.1% to 86.3%.