SRGAN for super-resolving low-resolution food images

Single image super-resolution, especially SRGAN, can generate photorealistic images from down-sampled images. However, it is difficult to super-resolve originally low resolution images that contain some artifacts and were taken many years ago. In this paper, we focus on the food domain because it is useful for our recipe-based web 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 down-sampled ones, we propose two approaches. The first one is a down-sampling method using noise injection to create desirable low-resolution images from high-resolution ones for model training. The second one is to train models for each target domain: we use the beef, bread, chicken and pound cake categories in our experiments. We also propose a novel evaluation method, Xception score. Compared with existing methods using qualitative and quantitative experiments, we find the proposed methods can generate more realistic super-resolved images.