A Scratch Detection Method Based on Deep Learning and Image Segmentation

With the improvement of product surface quality requirements in industrial production, machine vision has gradually become an important nondestructive testing method in the field of scratch detection. The traditional scratch detection method based on manually designed feature is susceptible to noise interference. Although the deep learning-based scratch detection method boasts strong robustness, it is difficult to completely and accurately segment the scratch through this method. We, therefore, propose a scratch detection method combining deep learning and image segmentation algorithm to realize recognition and segmentation of scratches with low contrast and small size. To effectively identify scratches, a multifeature fusion module was added on the basis of deep learning network framework. This module was designed according to the morphological characteristics of scratches. A principal component growth segmentation algorithm was designed for the extracted scratch prediction frame, and the scratch pixels were accurately segmented while the background noise was effectively suppressed. In the three scratch datasets under different application scenarios, the scratch recognition network proposed in this article has higher accuracy than the current mainstream target recognition methods when ensuring faster detection speed, and the segmentation results combined with the proposed principal component growth algorithm are more desirable than the current mainstream image segmentation methods.