Deep learning-based defect detection for mechanical parts

Abstract

To address challenges in mechanical component defect inspection, such as slender cracks, sparse defects, strong reflections on metal surfaces, and complex textured backgrounds, this study proposes an efficient visual defect detection model based on YOLOv10-SCI. The model uses a Shape-Sensitive IoU loss to improve bounding box regression for curved cracks and edge wear. It also adds the Convolutional Block Attention Module to enhance key defect features in both channel and spatial dimensions. In addition, it replaces stride-based convolutions with Space-to-Depth Convolution. Together with the lightweight StarNet structure, these changes reduce the number of parameters and help preserve fine details.Experiments show that YOLOv10-SCI achieves 91.41% Precision, 88.76% Recall, 90.06% F1, and 94.42% mAP. Compared with the original YOLOv10, this is an improvement of 8.13%, 5.05%, 6.57%, and 4.11%, respectively. The improved model shows high accuracy, strong robustness, and good lightweight performance in multi-class and multi-scale industrial defect detection. It is practical for engineering use and has strong potential for deployment. It also provides solid technical support for automated surface defect inspection in manufacturing.