Eigentextures: An SVD approach to automated paper classification

Eigentextures represent a SVD-based approach to texture classification which can be used in a trained or untrained setting. The method is analyzed by finding a concise relationship between the number of classes and the probability that the correct class will be selected, in terms of the number of comparisons that must be made. Because the method is computationally intensive, a simplified iterative version is suggested that can retain much of the classification power while reducing the computational burden. The advantages and disadvantages of these procedures are investigated in the context of the Historic Photo Paper Classification dataset. One feature of the eigentexture algorithms is that there is an inherent way to estimate the quality of the classification and to locate useful values of the parameters with or without training data.