Stamping part surface crack detection based on machine vision

Abstract

Cracks are the most common defects on the surface of stamping parts. Given the complex and varied structures of automotive stamping parts and their highly reflective surfaces, the current state-of-the-art methods lack effective automated inspection systems capable of precise online detection. Consequently, the identification of surface cracks in stamping parts on active stamping lines predominantly relies on manual visual inspection. However, this method is subjective, inefficient, and insufficient to meet the requirements for higher accuracy and detection rates. Therefore, this paper proposes a stamping part surface crack detection system based on machine vision. By devising an image acquisition module, high-resolution images of stamping parts are captured. Considering the strong reflectivity of the stamping part surface, an innovative gray-based contrast enhancement algorithm is proposed to adaptively balance the image contrast by comparing the grayscale values of the local window with the global image. To precisely locate and detect cracks, we design a novel crack online detection network (COD-Net), which is based on YOLOv9 as the backbone to improve detection efficiency and accuracy. This network incorporates the multi-scale crack attention (MCA) mechanism to obtain richer semantic information and more accurate feature representation. Notably, the crack detection context decoupling (CDCD) head is exploited in the detection head to improve localization accuracy and convergence speed. Moreover, we propose the CD-Loss, which introduces $\alpha$-WIoU in the detection box loss function to enhance model performance and speed up convergence. Our method significantly improves recall and achieves an [email protected] of 99.3% on the test set compared with other state-of-the-art methods. Furthermore, our detection system has been successfully applied to an automotive stamping lines.