Abstract: In order to enhance defect recognition in time-of-flight diffraction (TOFD) scan images for circumferential welds of pipelines under conditions such as background signal interference and uneven sample distribution, an improved YOLOv5s network model was proposed. In view of the defects with irregular shapes in TOFD images for circumferential welds of pipelines, deformable convolutions were introduced, enabling the network to adapt to the shape characteristics of defects and improving the feature extraction capability for irregular defects in TOFD images. To study the influence of interference waveforms such as direct wave and bottom wave on defect recognition in TOFD images, self-attention mechanisms were incorporated at different depths of the network, guiding the network to focus on subtle defect features while effectively suppressing the impact of interface waves on defect recognition. Furthermore, to tackle the issue of class imbalance in real-world defect samples, the SlideLoss loss function was employed to replace the original loss function, thereby enhancing the recognition accuracy for crack-type defects with limited sample sizes. Comparative experiments demonstrate that the improved network effectively suppresses complex background interference in TOFD images and improves defect recognition efficiency under imbalanced sample conditions. Compared to those of the original network, the overall mean average precision (mAP) and the average precision (AP) for crack-type defects have increased by 8.2% and 7.3%, respectively.