A small-object-sensitive representation learning approach for enhancing weld defect detection

Accurate detection of weld defects in images is crucial for ensuring the performance and safety of welded components. Existing methods face several challenges, including poor image quality, limited weld defect data, and difficulties in detecting weld defects of varying scales, especially small defects. To address these issues, we propose WDD-Net, a small-object-sensitive representation learning approach for weld defect detection. WDD-Net incorporates three key innovations: we propose a small-object-sensitive network that enhances the model’s ability to detect small objects and blurry features; we introduce contrastive learning to tackle the challenge of limited weld defect data; we propose an improved supervised loss function that balances the detection of both large and small weld defects. Extensive experiments on a weld defect dataset demonstrate that WDD-Net significantly outperforms state-of-the-art object detection models, improving air-hole mAP \(_{50}\) by 5.8% over the YOLOv9c baseline, from 91.7 to 97.5%.