Recognition of Handwritten Digits Based on Images Spectrum Decomposition

Abstract

Recognition of handwritten digits by convolutional neural network (CNN) using Fourier transforms of images as a preprocessing is considered. An algorithm of image preprocessing for effective CNN training and handwritten digits recognition is proposed. A discrete two-dimensional Fourier transform is applied to the original images. The real and imaginary parts are separated from the obtained complex values, as well as the amplitude and phase are calculated. Convolutional neural network is trained on the resulting characteristics obtained after Fourier transform. The proposed approach is tested on the MNIST database. The effects of image preprocessing using spectral decomposition and application of obtained different essential characteristics on the errors of handwritten digits recognition are estimated.