Texton Finders Based on Gaussian Curvature Of Correlation With An Application To Rapid Texture Classification
Shigeru Ando · 2005
This paper proposes a set of local image operators which are written by various combinations of second derivatives of a grayness correlation function. These operators, e.g., an edge finder, a corner finder, and an orientation finder, have the following excellent properties: 1) they are dimensionless thus virtually invariant to image contrast and resolution, 2) they respond exclusively to omnidirectional and unidirectional patterns thus can readily discriminate corners from edges, intersections from lines, dots from stripes, etc-, and 3) they can be computed regularly and paralelly thus are suitable for machine implementation of the operators. The performance of these operators is examined and compared with that of well known local operators. Demonstration of texton extraction and segmentation will be presented by using Julesz's patterns and a natural scene.