Zoom in on the target network for the prediction of defective images and welding defects' location

Automatic welding defects detection is crucial in intelligent welding manufacturing. However, the small size of defects hampers the advancement of automatic welding defects detection. This study proposes a Zoom in on the Target (ZIOT) network, which systematically performs tasks such as welded joint segmentation, defective image detection, and prediction of welding defect locations. The proposed model achieves 100 % recall and precision for segmenting the welded-joint region, surpassing the performance of the Otsu-based methods. The five-fold cross-validation experiments indicate the proposed model can distinguish defective and non-defective X-ray images with an accuracy of 98.4 %. The segmentation of welded joints contributes to a 10 % improvement in the average precision of predicting the location of welding defects. Moreover, the ZIOT network demonstrates superior performance when compared to classical models, including Faster R-CNN, YOLO, and Swin Transformer. The ZIOT network exhibits significant potential for application in detecting welding defects within X-ray images acquired through the DWDI technique.