Multi-scale Fusion for Arbitrary-scale Image Super-resolution

Abstract

Convolutional neural network (CNN)-based single image super-resolution (SR) methods have achieved superior performance on some discrete scale factors. Majority CNN-base methods only work for some integer scale factors, such as 2, 3, and 4. Meta-SR method manages to work for more discrete scale factors, such as 1.1, 1.2, etc. However, the scale factors for SR should be arbitrary in practical applications. To extend CNN-based SR methods to an arbitrary scale, in this paper, we propose a multi-scale fusion method for arbitrary-scale SR (MSFASR). In our MSFASR, we first input a low-resolution (LR) image into an existing SR model to generate two different high-resolution (HR) images. Then we downscale/upscale HR images to predict the SR results with target resolution. By exploring scale factor preference, we combine the SR results to get the final SR image. Experimental results with Meta-SR method demonstrated that the MSFASR method could extend CNN-based SR method from discrete to continuous scale and achieve good SR performance on arbitrary scales.