Research on internal defect detection of metal components based on deep learning

Abstract

To address the challenges of internal defect detection in metal components, including large variations in defect scale, irregular defect morphology, grayscale similarity between defects and the matrix, and interference from complex material backgrounds, this study proposes an efficient detection model for internal defects in metal components based on A YOLOv10 RM. MPDIoU is introduced to optimize the loss function and improve bounding box regression accuracy. AssemFormer is used to enhance the extraction of multiscale defect features under complex material backgrounds, while the Rectangular Self Calibrated Module is incorporated to strengthen spatial location modeling of defects and global structural perception. The results show that the improved A YOLOv10 RM network achieved a Precision of 96.54%, a Recall of 96.09%, an F1 score of 96.31%, and an mAP of 98.92%, representing improvements of 2.86%, 3.29%, 3.04%, and 2.61%, respectively, compared with the original YOLOv10 network. The improved model demonstrates excellent detection accuracy, robustness, and generalization capability in intelligent internal defect detection of metal components. It also shows strong engineering applicability and potential for broader deployment, providing solid technical support for quality assurance of high end equipment and nondestructive testing in intelligent manufacturing.