Abstract: A welding defect detection algorithm YOLO-SBRS based on YOLOv8n was proposed for welding defects with characteristics of multiple scales, complex shapes, and vulnerability to background interference. First, spatial and channel reconstruction convolution (SCConv) was used to improve the C2f module of the backbone network. At the same time, a spatial pyramid pooling fast average pooling with biformer attention module (SPPF_ABF) was designed, which replaced the maximum pooling of the original SPPF module with the average pooling operation, and a two-layer routing transformer attention mechanism was introduced. Secondly, a reparameterized generalized feature pyramid network (RepGFPN) was used to optimize the feature fusion part. Finally, based on the principle of parameter sharing and the introduction of a module combining space-to-depth and non-strided convolutional layers (SPD_Conv), the detection head was improved to achieve light weight and improve the detection capability of the network for complex defects. Experimental results show that the precision of the improved algorithm and mean average precision at 50% intersection over union (mAP50) are improved by 3.1% and 2.8%, respectively, which provide an efficient and feasible solution for welding defect detection.