Convolution Using Discrete Cosine Transforms for Improving Performance of Convolutional Neural Networks

Abstract

Convolutional neural networks (CNNs) are widely used in many areas. They feature convolutional layers that focus on spatial local node connections rather than full node connections. This makes networks much more efficient for spatial information. The convolution is a mathematical operation on two functions and can be calculated using the discrete Fourier transform (DFT). Due to the close relation to the DFT, the discrete cosine transforms (DCTs) can be used for the calculation. In this paper, we focus on the convolution using DCTs for improvement of the performance of CNNs. The periodicity and symmetry inherent in the DCTs generate larger output feature maps. The proposed method in simple CNNs is demonstrated and the efficacy of the proposed method is testified using CIFAR-10 dataset.