High precision small defect identification network framework with enhanced edge feature perception for arc welding

Non-destructive testing (NDT) targeting small defects within arc welding seams is critical for ensuring the reliability of advanced manufactured products. These small defects present significant challenges for detection, as they often exhibit visual patterns similar to the background, making them difficult to distinguish using conventional methods. To address this problem, a detection framework is proposed for small defect identification. First, a Multi-Scale Edge-Enhanced Attention (MSEA) module is designed. It specifically focuses on edge and texture signals of small defects with novel explicit foreground guidance. Second, to mitigate feature coupling between defect and weld characteristics, Ghost-Shuffle Convolution (GSConv) is embedded to enhance low-level image feature extraction. Finally, a Scale-Based Dynamic Loss (SDLoss) is designed that adaptively enlarges the localisation-error gradient for small targets according to the pixel area of their candidate boxes, ensuring that small defects receive sufficient attention during training. Extensive experiments demonstrate that the proposed model achieves the best mean Average Precision (mAP) of 82.2% on an arc welding defect dataset, and also attains superior performance on two public benchmarks, significantly outperforming several state-of-the-art (SOTA) methods.