Aiming at the problems of low efficiency and insufficient accuracy of automatic welding seam inspection in industrial scenes, WDFCOS model for weld seam defects was proposed. Based on FCOS model, a lightweight backbone network, RelementNet, was designed to extract image features more efficiently. PAN(pyramid attention network) structure was used to strengthen the original feature fusion network and form a bidirectional feature fusion network. A new multi-scale attention CGM module was designed to reduce the information decay in the process of feature fusion. Through experiments on self-built data sets, AP(average precision) can reach 82.6% by WDFCOS model, which is 6.9% higher than FCOS model. At the same time, the detection rate can reach 49.1 frames per second on the edge equipment more suitable for industrial scenarios, and the detection time is reduced by 84.4% compared with the baseline model, achieving a faster and more accurate weld defect detection.