This study proposes a weld defect detection method that addresses the low recognition accuracy observed in industrial weld surface defect detection tasks owing to the small target size and high similarity to the background. The proposed weld defect detection method is based on improved YOLOv9. First, a dynamic sparse attention module is introduced to improve the feature extraction network of the main branch. This module enables the ability of the main branch network to extract features of targets and backgrounds in complex backgrounds via the long-distance content awareness of self-attention mechanisms and flexible computation allocation. Subsequently, an adaptive dynamic convolution module is incorporated into the backbone network of the auxiliary branch, which dynamically adjusts the convolution position based on learned offsets to accurately locate and fit target features, thereby improving the detection performance for small targets. Finally, an MPDIoU function is proposed to improve the loss function, maximizing the overlap between predicted bounding boxes and true bounding boxes, which enhances the accuracy of weld defect detection and the convergence speed of the network. Results of the experiments conducted on the self-constructed weld defect data set, WELD-DETECT, show that compared to the baseline network YOLOv9, the proposed DA-YOLO network achieves 4.9 percentage points improvement in detection accuracy and 5.5 frame/s increase in detection speed.