Abstract: Digital Radiography (DR), with its high resolution, real-time imaging capabilities, and flexible post-processing advantages, has developed into an important non-destructive testing method in weld inspection. AI-based intelligent defect detection technology has emerged as a key development direction in DR image processing, notably enhancing detection efficiency across various application scenarios. However, the significant size variation among different weld defects remains a challenge for existing intelligent defect recognition methods. To address this, this paper proposes an EP-YOLO model based on efficient feature extraction and precise defect localization. The backbone network of the model utilizes a compound scaling strategy and parameter allocation optimization to achieve efficient multi-scale feature extraction for defect detection. Meanwhile, the loss function introduces a target size-adaptive penalty factor and incorporates a gradient-adjusting strategy based on anchor box quality, effectively improving the localization accuracy of defect objects and accelerating model convergence speed. Experimental results demonstrate that the EP-YOLO model exhibits outstanding performance, achieving a mAP@50 of 93.1% in precise localization and classification tasks and reaching a mRecall of 91%, effectively reducing defect misses. Compared to mainstream YOLO series algorithms, it shows significant advantages. Notably, the model excels in detecting small-scale pore defects, with AP@50 surpassing the best-performing mainstream industrial detection model by 8.2% and the recall exceeding 7.8%. Furthermore, the EP-YOLO model has only 4.6M parameters and a computational complexity of 8.4 GFLOPs, achieving detection speeds comparable to mainstream YOLO series models, with a frame rate of up to 70 FPS. This model combines high accuracy and fast detection speed, providing strong technical support for the practical application and deployment of intelligent weld defect detection.