Abstract: AI-based weld digital radiography (DR) image defect segmentation technology can effectively assist in defect grading. However, existing technologies still face significant challenges in segmentation accuracy when dealing with low-contrast and small-sized defects. To address this, this paper proposes an intelligent defect segmentation network, EA-Unet, based on efficient feature extraction and an adaptive sparse attention mechanism. The backbone of the model employs a compound scaling strategy to optimize depth, width, and resolution, enabling efficient extraction of multi-scale features for weld defects. Additionally, an adaptive sparse Transformer block is introduced to mitigate noisy interactions of irrelevant areas and reduce feature redundancy in the channel domain. Furthermore, transfer learning techniques are employed to alleviate the issue of insufficient samples in weld DR image defect segmentation. The results show that, compared with mainstream industrial segmentation models, the proposed EA-Unet model achieves a 5.7% improvement in mIoU and a 4.9% improvement in mPA, demonstrating improved segmentation accuracy and pixel-level classification capabilities. Additionally, the EA-Unet model has 16.4×10⁶ parameters and a computational complexity of 207 GFLOPs, with a detection frame rate of 53.0 FPS. With its excellent segmentation performance, the EA-Unet model provides strong technical support for the practical application and deployment of weld DR image defect segmentation.