Visual perception of surface roughness via maximum variance direction of image patches

Abstract

Most existing visual roughness perception algorithms are pixel-wised operations, which lead to unstable roughness measurements. To address this problem, a novel visual roughness perception in the maximum variance direction of image patches is proposed in this article. Specifically, each patch is assumed to be a point in a multi-dimensional Euclidean space. To improve the signal-to-noise ratio of roughness, all points are projected onto its maximum variance direction based on principal component analysis. In this direction, the variance of these points is characterized as roughness. A new roughness metric dataset is constructed by scanned images of sandpapers with different mesh sizes to verify the effectiveness of the proposed algorithm. Experiments on the dataset show the proposed method outperforms many computer vision-based methods, such as the gray-level co-occurrence matrix approach and the approach.