Abstract: To solve the problem of robot adaptive weld grinding, this paper proposes a weld detection algorithm based on improved YOLOv5, which realizes the identification and initial positioning of the welds. The welding robot is used to produce weld seams with different parameters and shapes, and a dataset containing 3 996 weld images is self-made for deep learning. The YOLOv5s model is selected for training, and the GAM attention mechanism module is added to the backbone. At the same time, GhostNet is introduced, and the Conv module and C3 module of the original model are respectively replaced by the GhostConv module and C3 Ghost module. The mAP@0.5 of the improved YOLOv5s-GhostNet-GAM model reaches 90.21%, 4.05% higher than that of the original YOLOv5s model. At the same time, the number of parameters is reduced by 5.64%, and the FLOPs are reduced by 27.44%. The detection rate is 23.47 FPS. It meets the requirements of robot adaptive weld grinding for identification accuracy and later software deployment.