Synthesis of data-dependent filters for digital image processing and recognition

Linear filtering remains an attractive method of image recognition because of simplicity of its performance in standard computing systems. Moreover, even nonlinear procedures of image processing include, as a rule, linear filtering for preliminary selection and transformation of classifying signs and compression of processed information to acceptable values. The efficiency of linear filtering is defined mainly by the proper choice of filter set, that must correspond to the peculiarities of the problem. Generally, it is desired that the total number of filters used would be as small as possible. It seems promising to use filters, constructed for some optimum criterion which take into account the statistical properties of the general image population. A well known example of such an approach is given by the Karhunen-Loeve decomposition. It is evident, that KL-decomposition is optimal for representation of the images, not for their discrimination. We suggest the procedure of discriminant filters construction, based on an a priori classified training set, and present some results of its practical application