Abstract: detection，an improved YOLOv8 detection algorithm integrating coordinate attention and asymptotic feature pyramid network is proposed. In the data augmentation stage，a combination of Mixup and random rotation is employed to alleviate the problem of insufficient samples of rare defect categories. For network improvement，certain modules in the cross-stage sections of the YOLOv8 backbone network are replaced with coordinate attention-based modules to enhance the sensitivity to defect location. Additionally， the detection head is replaced with a asymptotic feature pyramid network head to improve multi-scale feature fusion. Subsequently， comparative experiments are conducted using the open-source welding defect dataset GC10-DET. The experimental results show that the proposed method achieves improvements in both detection accuracy and recall rate for small defects，providing a high-precision and high-robustness solution for welding component defect detection.