Kernel Symmetry for Convolution Neural Networks

Abstract

A convolution neural network (CNN) uses kernels to filter applied images. These kernels learn their coefficients' values during the training process, thus they do not possess any centrosymmetry. Hence, the phase responses for these kernels are neither zero-phase nor linear-phase. This technique adds a group delay distortion to the filtered images. In this paper, we constrain the values of the kernels' coefficients to be centrosymmetric. This scheme guarantees the prevention of any distortion in the filtered images. In the proposed method, the CNN trains all the kernel coefficients as normal. Then every two-centrosymmetric coefficients are set to their average. This does not affect much the accuracy of the CNN. The proposed algorithm may be used to improve images generated using generative adversarial networks (GAN), autoencoders, image segmentation, and all other algorithms that generate images or video using CNN. This point still requires further study.