A weld defect detection method based on large convolutional inverted-mixed residual networks with unbalanced samples

Abstract In the intelligent detection of welding defects, the sample set is highly imbalanced, with many more normal samples than abnormal ones. However, the existing methods have a weak ability to extract global features in the backbone part, and are prone to favor most types of samples under unbalanced sample conditions. There is still room for further improvement of the existing target detection accuracy. To address these issues, a welding defect detection method is proposed based on a large convolutional inverted-mixed residual network under imbalanced samples. First, UniRepLKNet and SPPF (spatial pyramid pooling and feature fusion) are integrated into the backbone to form the SPPF-UniRepLK module. Second, a large convolutional kernel is introduced in this module to enhance the focus on global features and to capture multi-scale welding defect information better. The inverted-mixed residual mobile block module is designed by combining the inverted residual structure with an attention mechanism, thereby improving the model’s feature extraction for abnormal samples. Third, the Focal Scale-Invariant IoU loss function is developed to adjust category weights dynamically, enhancing the model’s ability to detect abnormal samples. Experimental results demonstrate that the proposed method achieves a precision of 93.2%, a recall of 86.0%, and an mean average precision (mAP) of 89.5% in welding defect detection tasks, surpassing the current state-of-the-art model, GhostNetV2, by margins of 2.2% in precision and 0.9% in mAP, while maintaining comparable recall performance.