C-ESRGAN: Synthesis of super-resolution images by image classification

Abstract

With the development of deep learning, super-resolution image synthesis techniques for enhancing low-resolution images have advanced remarkably. However, mainstream algorithms focus on improving the quality of the entire image on average and this may result in blurring. In this paper, we propose three key components for synthesizing super-resolution images that can reflect the fine details of an image. We synthesize super-resolution images by image classification. First, the neural network weights learned using the images in the same image category were utilized in synthesizing super-resolution images. For this purpose, image classification was performed using a transfer-trained ResNet. Second, SENet was applied to the generators in our proposed method to obtain detailed information about the images. Finally, the feature extraction network was changed from VGG to ResNet in order to get more important features. As a result, we achieved better image evaluation values (PSNR, NIQE) for the super-resolution images of dogs and cats compared to the previous studies. Furthermore, the images were generated more naturally on the benchmark dataset.