Abstract: To enhance the reliability of γ-ray radiographic film defect detection of welds, a defect detection model based on improved YOLOv5s was proposed, achieving efficient and accurate detection of radiographic film defects of welds in complex environments. In response to excessive channels and redundant information in the convolutional network of the original YOLOv5s model, firstly, the SCCONV network module was integrated into the C3 module of the backbone network, reducing redundant information and boosting the detection performance of the model. Secondly, in light of the characteristics of weld defects, such as morphological diversity, scale variation, and low contrast, convolutional attention modules (CBAM and SE) were introduced to enhance the attention of the model to the region of interest. Finally, the EIoU loss function was employed to replace the CIoU loss function in the traditional YOLOv5s in the bounding box regression detection, significantly improving detection precision and robustness of the model. Experimental results demonstrate that the improved model has significant improvements in the indices such as precision and recall rate, compared with the traditional YOLOv5s algorithm. They can be manifested in the fact that the accuracy is increased by 4.2%; the recall rate is increased by 3.2%; the mean average precision is increased by 3.4%. The effectiveness of the method in the γ-ray defect detection of welds was verified.