An Efficient CNN-Based Detector for Weld Defect Detection Based on Enhanced YOLOv8

The efficient and accurate detection of weld defects is of great significance in the industrial process, which can effectively ensure the safety and reliability of welded structures. In recent years, with the rapid development of deep learning, the You Only Look Once (YOLO) series of object detection models have also achieved great success due to their high efficiency and accuracy. However, the detection of weld defects still faces many challenges, such as diverse defect categories, large size differences, and poor texture. To address the above issues, an efficient CNN-based detection model based on YOLOv8 is presented for weld defect detection. Specifically, an enhanced backbone is built with dynamic snake convolutions (DSConv) to construct a feature perception unit with topological deformation capabilities. In addition, in order to effectively extract multi-scale features, a Large Selective Kernel Network (LSKNet) is introduced in the feature fusion stage, which can better integrate multi-scale features and further improve the model’s detection ability for different sizes of weld defects. Finally, to further improve the overall performance of the weld defect model, a Wise-IoU loss function is adopted as the regression loss function of the detection model to optimize the model’s positioning accuracy on micro weld defects. Experiments show that proposed model has achieved good detection performance on the public weld defect dataset (LoHi-WELD), demonstrating its effectiveness in industrial scenarios.