A More Focus on Multi-degradation Method for Single Image Super-Resolution

Abstract

Single Image Super-resolution (SISR) aims at reconstructing a High-Resolution (HR) image from a Low-Resolution (LR) one. Recent works, especially in the deep learning-based approach, mainly define and resolve the problems of LR images degraded by a fixed degradation kernel, typically bicubic interpolation. However, this assumption can hardly be practical since an input image may suffer from many other deteriorations (e.g. blur or noise). Previous works tackle such multi-degradations by proposing new models, targeting at lessening the restrictions of learning-based method and taking advantages of CNN architecture. Unfortunately, they ignore the existing state-of-the-art CNN-based SISR models that are trained on a fixed degradation kernel. In this work, we introduce a context-extending module that generates on-the-fly more realistic types of degradation. We also come up with a comprehensive cross-degradation loss function enabling the model to better adapt real-world conditions. With this proposal, we can generalize arbitrary end-to-end learning-based networks. Evaluating by Peak Signal-to-Noise Ratio (PSNR) metric, our proposed method outperforms the EDSR baseline a significant amount of 34.5% (from 20.14dB to 27.09dB) on noisy images meanwhile sustaining the comparable results on the bicubic downsampling factor.