Image Super-resolution Based on Sparse Representation With Joint Constraints
Jie Xu, Cheng Dan Deng, Xianglong Liu, Jie Li · 2014
Learning-based image super-resolution (SR) methods are prone to introduce artifacts into resultant high-resolution (HR) images, while reconstruction-based ones tend to blur fine-grained parts and result in unnatural results. To solve these problems, this paper proposes a novel image SR algorithm based on sparse representation with joint constraints. The self-similar redundancy structure in input LR image is learned to construct sparse representation dictionary. Sparse representation, gradient histogram preservation (GHP) and non-local means (NLMs) are incorporated into a maximum a posterior (MAP) estimation framework. Then, we gradually magnify the LR input in the current frame to the desired size and obtain a high-quality SR image. Experimental results demonstrate that the proposed method outperforms the state-of-the-arts.