Multi-frame Super-resolution via Generative Image Model

Abstract

Recent start-of-the-art video super-resolution (VSR) network has achieved huge success on popular video datasets. However, these methods may fail in real-word video data, since low-resolution (LR) images are generated by bicubic interpolation on corresponding high-resolution (HR) images, which are not subjected to the real-word degrading model. Besides, the training dataset may lack sub-pixel motion between frames, which is crucial for image detail restoration. To address this issue, a novel generative model is proposed for constructing low-resolution image sequences with sub-pixel motion information and various blur kernel. This data augmentation is an effective way to improve the performance of dealing with spatial-temporal information. Based on this strategy, the super-resolution results are more reliable and robust. Extensive experiment shows our proposal helps the VSR network make full use of sub-pixel information to reconstruct a reliable high-resolution image with rich details.