Automated Metal Surface Defect Detection

Abstract

Product defect detection is one of the essential steps in quality control to ensure product safety. Inspection needs to be conducted regardless of during or at the end of the manufacturing process to ensure all products complies with specification and requirements. In-process defect detection is conducted to identify any deviations or defects in the product to ensure all pieces are consistent and safe to use. However, manual quality inspection procedures are often time-consuming, expensive, and prone to errors especially for manufacturers that conduct large-scale production on a daily basis. Hence, many industries have started to leverage and incorporate technologies such as IoT devices, Computer Vision, Artificial Intelligence and Deep Learning in the manufacturing process for a more robust and efficient defect detection system. This project aims to identify the most suitable image segmentation method for patch defect and scratch defect respectively and accurately localize the patch defect on metal surfaces and evaluated using Intersection over Union (IoU). In this project, a total of 3 image segmentation methods are attempted namely threshold-based segmentation, edge-based segmentation and clustering techniques are implemented and compared for detecting patch and scratch defects. We successfully identified that the threshold-based segmentation method is more suitable for patch defects whereas the edge-based segmentation method is more suitable for scratch defects. Among the attempted threshold-based segmentation, namely simple thresholding, adaptive thresholding and Otsu’s Binarization, we discovered that the best technique to detect patch defects is Otsu’s Binarization.