Surface Defect Detection Using Singular Value Decomposition

Abstract

The defect inspection on the surface becomes a critical task in industrial manufacturing. Defects often appear on surfaces of steel, plastic, and glass. There are a lot of research efforts to develop advanced image processing methods to improve defect detection. Based on the assumption that each defect image could be decomposed into two components: the defect-free background component and the defect foreground component. The background reflects the similarities of different regions, and the foreground reflects unique defect information. In this work, we propose the singular value decomposition-based (SVD) algorithm for color images to detect surface defects. First, we determine the residual component by using the SVD-based full rank approximation. Next, we recognize the structural part by choosing the suitable matrix rank for the SVD-base structure-rank approximation. The addition of the residual part and the structural part becomes the background image. The result of subtraction between the original image and the background image carries the defect information. Finally, we locate the rectangle boundary that surrounding the defect based on the simple thresholding operation.