Abstract: Traditional manual film evaluation and ultrasonic phased array techniques for detecting girth weld defects in oil and gas pipelines suffer from low efficiency, subjectivity, and missed minor defects. In this study, an intelligent recognition model was developed using the Easy Deep Learning (EasyDL) platform and You Only Look Once version 10 (YOLOv10) algorithm to enhance defect recognition accuracy and efficiency. A dataset of 1 000 radiographic films featuring five defect types—slag inclusion, porosity, cracking, incomplete penetration, and lack of fusion—was compiled, incorporating real-world engineering noise and imaging interference. Leveraging EasyDL's automated annotation, pre-trained models, and versatile deployment, rapid training and iteration were achieved. The EasyDL model attained an overall mean Average Precision (mAP) of 83.1%, recall rate of 77.2%, and over 85% recognition rate for defects such as cracks. In contrast, YOLOv10, with its lightweight design and architecture without Non-Maximum Suppression (NMS), achieved an mAP@0.5 of 70.5%, precision of 81.4%, and recall rate of 56.9% on the same dataset. The comparison results indicated EasyDL offered superior engineering adaptability and rapid deployment, whereas YOLOv10's performance was constrained by the need for professional parameter tuning. This study pioneers the application of NMS-free object detection and zero-threshold AI platforms in pipeline girth weld defect recognition, reducing reliance on manual film evaluation and multimodal preprocessing. Future work should focus on cross-platform model integration and noise-augmented training to enhance sensitivity to minor defects, thereby advancing the standardization and workforce efficiency of intelligent operation and maintenance for oil and gas pipelines.