Local Linear Transforms for Texture Measurements
Signal Processing, vol. 11, no. 1, pp. 61–79, July 1986.
The Nth order probability density function for pixels in a restricted neighborhood may be characterized by a set of N histograms (or some corresponding moments) computed along appropriately chosen axes. The projections on those axes are obtained from a local linear transform of the local neighborhood vector. This approach is closely related to filter bank analysis methods and gives a statistical justification for the extraction of texture properties by means of convolution operators or local matches. Optimal and sub-optimal linear operators are derived for texture analysis and classification. Experimental results indicate that the method is robust, flexible, and that it performs as well as standard co-occurrence based methods for texture classification. The proposed approach enables texture characterization with a lower number of features and it is also computationally more appealing.
@ARTICLE(http://bigwww.epfl.ch/publications/unser8602.html, AUTHOR="Unser, M.", TITLE="Local Linear Transforms for Texture Measurements", JOURNAL="Signal Processing", YEAR="1986", volume="11", number="1", pages="61--79", month="July", note="")