High Performance Example-Based Image Super-Resolution

Super-resolution (SR) image reconstruction is a typical inverse problem in computer vision. With the flourishing of digital image and video, SR has attracted more and more research attention due to its extensive applications in real world. Especially, there are many occasions that we want to increase the resolution of a single image, and the classical multi-frame SR techniques are not applicable. Single-image SR is a severely ill-posed problem, the solution to which thus requires effective prior information as a supplement. Take advantage of the correspondence between co-occurrence image patches at two difference resolutions, recently developed example-based SR has been demonstrated a breakthrough and brings significant performance improvement in terms of image reconstruction quality over previous methods. However, the limited mapping accuracy from low-resolution (LR) to high-resolution (HR) is a big barrier to further exploit the potential of example-based SR.In this paper we propose three different methods to improve the mapping accuracy in example-based SR. First, a feature enhancement scheme, which combines prefiltering integrated interpolation and non-blind sparse prior deblurring, is utilized to reduce the intrinsic dimensionality gap between HR and LR spaces and thus the ambiguity during feature mapping. Second, soft information and decision is employed to depict the one-to-many correspondence between LR and HR patches by passing explicit parametric distributions on Tanner graph, boosting the flexibility and reliability of training and learning in example-based SR. Third, a robust SR framework, which combines adaptive PDE regularization and example-based SR, is developed to simultaneously improve the resolution and perceptual quality of widely distributed web image/video degraded by both downsampling and compression.Extensive experiments have been conducted on a variety of images (as well as videos if applicable) to evaluate the above proposed methods. Distinct performance improvement in terms of both objective and subjective image reconstruction quality demonstrates the effectiveness of our methods compared with state-of-the-art works.