Automatic construction of single frame super-resolution using Cartesian Genetic Programming

Abstract

In this paper, we propose a single-frame Super-Resolution (SR) method using Cartesian Genetic Programming (CGP). Our method is to learn relationship of pixel values between high-resolution (HR) image and low-resolution (LR) image using CGP, and we construct a SR rule of generating SR image from a LR input image. A single pixel and its neighbor pixels of the LR input image are set to the inputs of CGP. And then, pixel values of the SR image are obtained from the calculated outputs of CGP. Therefore, the SR image is generated from the LR input image. In addition, multiple CGP can improve the quality of SR image. Because our method is to perform for each pixel independently, our method is suitable to parallel processing. Therefore, in order to reduce computational cost, we use parallel processing with graphics processing unit (GPU). Experimental results show efficient processing is constructed. Our method is little less quality than one conventional work which is the state of the art method on image quality, however to perform overwhelmingly faster than the conventional work. We can construct fast and accurate single-frame super-resolution.