Research On Single Image Super-resolution Reconstruction

Single image super-resolution aims to reconstruct an HR image of the same scene using only one LR input image combined with specific prior knowledge, in order to overcome the inherent resolution limitations of imaging system. In recent years, the learning-based SR methods have become research hotspots gradually. Based on the study of related algorithms of sparse representation and dictionary learning, this thesis proposes two single image SR methods as below:Because of the low learning accuracy of dictionaries in the sparse representation based SR method, semi-coupled dictionary learning(SCDL) framework which relaxes sparse representation invariance assumption is used to enhance learning ability. Then, a nonlocal similarity constraint of sparse domain is added in order to exploit geometry information and stabilize the sparse decomposition, and the resulting optimization is solved by the proposed modified Feature-Sign Search algorithm. Further, on the basis of initial clustering, the data is re-clustered by mapping error of sparse domain, and the whole training phase is performed by the proposed alternate heuristic learning framework. Besides, in reconstruction phase, this thesis suggests an alternate heuristic reconstruction scheme and an error compensation stage. Experimental results show that the proposed method achieves higher reconstruction quality and better visual performance.On the basis of aforementioned method, this thesis suggests uncoupled dictionary learning framework to reduce computational complexity. Then, restricted Boltzmann machine(RBM) is adopted to infer the Sparsity Pattern!, and under a criterion of minimizing reconstruction error, the dictionary learning is incorporated into statistical prediction model. Moreover, the basic scheme is extended by clustering based on sub-pixel sampling pattern and directional information. Experimental results show that the proposed method accelerates the SR reconstruction phase significantly, while achieves competitive reconstruction quality.