Low-complexity CNN with 1D and 2D filters for super-resolution

Abstract

This paper proposes a low-complexity convolutional neural network (CNN) for super-resolution (SR). The proposed deep-learning model for SR has two layers to deal with horizontal, vertical, and diagonal visual information. The front-end layer extracts the horizontal and vertical high-frequency signals using a CNN with one-dimensional (1D) filters. In the high-resolution image-restoration layer, the high-frequency signals in the diagonal directions are processed by additional two-dimensional (2D) filters. The proposed model consists of 1D and 2D filters, and as a result, we can reduce the computational complexity of the existing SR algorithms, with negligible visual loss. The computational complexity of the proposed algorithm is 71.37%, 61.82%, and 50.78% lower in CPU, TPU, and GPU than the very-deep SR (VDSR) algorithm, with a peak signal-to-noise ratio loss of 0.49 dB.