Implementation of an efficient super resolution algorithm for game images using sparse matrix multiplication

Abstract

We present a super resolution algorithm that increases the resolution of game scenes. Since game scenes are required to be rendered in real time, rendering high-resolutional game scenes in real time is very challenging. Therefore, increasing the resolution of game scenes rendered in low resolution is very promising. In order to present a fast super resolution algorithm, we implement our super resolution algorithm using a sparse matrix operation that can be accelerated through a matrix multiplication unit (MMU). Since many open source GPUs present an acceleration environment, our MMU-based approach can be implemented in many open source GPUs. As a baseline of our approach, we present a sparse matrix multiplication formula of bilinear interpolation and bicubic interpolation. We furthermore present a sparse matrix multiplication form of a convolution operation, which is necessary for deep learning-based super resolution algorithms. The filters employed in the convolution operations are converted to a series of sparse matrices, which are multiplied to complete the deep learning-based super resolution model. We demonstrate the performance of our approach by measuring peak signal-to-noise ratio (PSNR) and mean opinion score (MOS) of the result images generated by our approach and by the existing methods.