О повышении эффективности вычислений при классификации изображений

Abstract

Automatic classification of images using the Bayes classifier with integer coefficients is considered. It is shown that by representing the polynomials used in the classifier that differ from the generally accepted ones through indicating only nonzero coefficients, it is possible to save memory and reduce the number of operations while keeping the same classification quality. The polynomial coefficients are rounded and represented using integers (fixed-point numbers). A polynomial storing method that makes it possible to store only nonzero coefficients of the polynomial is analyzed. In using this method, the vectors and numbers used for carrying out classification are stored in the usual way. The MNIST database was used for carrying out the experiments. Seven bits were taken to represent the polynomial coefficients. For storing coefficients without rounding, 11.77 MB of memory is needed; for storing the rounded coefficients in the usual form, 2.93 MB of memory is needed; and for storing only nonzero rounded coefficients, 1.3 MB is needed. The number of addition and multiplication operations required to make a decision about the image class has been reduced significantly. An economical method for storing matrices with polynomial coefficients on a disk is presented. In the experiment with the MNIST database, only 0.35 MB of memory had to be used for storing the polynomials in the database. The study results can be used in elaborating software and hardware for automatic classification of images. It is also of interest to study similar techniques to improve the efficiency of computations in other classifiers.