Fourier transform-based image classification using complex-valued convolutional neural networks

Abstract

Although complex-valued convolutional neural networks (CVCNNs) have shown an improvement over their real-valued counterparts (RVCNNs) when trained on real-valued images, in order to harness the full potential of CVCNNs, they should be used with fully complex-valued inputs. Because every image has one and only one Fourier transform, the problem of classifying real-valued images is equivalent to the problem of classifying their complex-valued Fourier transforms. Experiments done using the MNIST, SVHN, and CIFAR-10 datasets show an improved performance of CVCNNs trained on Fourier-transformed images over CVCNNs trained on real-valued images (for which the imaginary part of the input is considered to be zero), and over RVCNNs.