Improving SRGAN for Super-Resolving Low Resolution Food Images

Abstract

Super resolution, especially SRGAN, can generate photo-realistic images from downsampled images. However, it is difficult to super-resolve originally low resolution images taken many years ago. In this paper we focus on food domains because it’s useful for our service if we can create better looking super-resolved images without losing content information. Based on the observation that SRGAN learns how to restore realistic high-resolution images from downsampled ones, we propose two approaches. The first one is downsampling methods using noise injections in order to create desirable low-resolution images from high-resolution ones for model training. The second one is training models for each target domain: we use {beef, bread, chicken, poundcake} domains in our experiments. Comparing to existing methods, we find the proposed methods can generate more realistic super-resolved images through qualitative and quantitative experiments.