Deep Feature Translation Network Guided by Combined Loss for Single Image Super-Resolution

Abstract

Single image super-resolution (SISR) which aims to infer a high-resolution (HR) image from a single low-resolution (LR) image has wide applications such as surveillance and medical image processing. However, existing methods which aiming at minimizing the mean squared error (MSE) always get high objective quality, i.e., peak signal-to-noise ratios (PSNR), but their results are blurry which lacks high-frequency details thus are perceptually unsatisfying. Some recently proposed Generative Adversarial Networks enhance the perceptual quality greatly, but their objective quality is very low, which means their generated texture details are not faithful to the real image. In this paper, we adopt a multi-scale HR construction process to generate HR images gradually to achieve large upscaling factors. For each level, the generation of HR difference features from LR features is taken as a feature translation process, and deep image feature translation network (DFTN) is designed. To recover finer texture details, we combine three loss functions: content loss, a novel fine-grained texture loss and adversarial loss in our model optimization. We desire that the content loss ensures the LR results faithful to the original image, and the other two losses push our model to capture the manifold of natural images. Experiments confirm that our model can achieve the state-of-the-art results in different evaluating metrics, including both objective and perceptual quality evaluations. Therefore, our method can generate HR images with fine texture details and faithful to original images.