Abstract
A set of image feature operators which are derived from curvatures of local correlation function, i.e., sums of products of grayness derivatives, is described and examined. First, a quadratic surface defined by the curvatures are related to local grayness properties which lead to nonlinear operators being sensitive to 1) overall variations, 2) equidirectional variations, 3) unidirectional variations, 4) orientation, and 5) maximum unidirectional variations. Then combining these operators, we obtain dimensionless and normalized expressions which are independent to contrast and resolution of image features. Basicly, they act as statistic operators to extract 6) overall, 7) equidirectional, and 8) unidirectional normalized variations of image textures. By functional analysis, however, they are shown to respond also in deterministic and exclusive manner to structural features such as blobs, edges, lines, corners, intersections, etc. Results show unique characteristics of the operators such as independence to clearness, normalized outputs, classified response on directivity, and stable orientation outputs.
