Research On Single-frame Image Super-resolution Reconstruction Algorithm Based On Instance Mapping Learning

Single image super resolution(SISR) algorithms take a low resolution(LR) image as input and reconstruct the corresponding high resolution(HR) image with specified magnification. SISR algorithms can be categorized into interpolation-based, reconstruction-based, and exemplar learning based methods which are superior to the others. In this paper, the fundamentals of exemplar-based SISR methods are analysed thoroughly and based upon which, the framework of learning exemplar mappings is outlined to enhance the capability of modeling feature spaces and complex nonlinear mapping among them. Under that framework, two SISR algorithms are proposed.As for sparse coding-based exemplar learning methods, a jointly trained dictionary pair in the training phase cannot guarantee co-occurrence of sparse representations of LR and HR features in the reconstruction phase when the LR dictionary individually represents the input LR feature. In order to solve the problem, the sparse domain selection(SDS) based SISR algorithm is proposed. As assumed in SDS that there is a deterministic mapping relationships between sparse representations of LR and HR features, then an optimal object follows. By integrating with quadratically constrained quadratic program, sparse coding, and ridge regression, we can obtain the optimized mapping, HR dictionary, and sparse representation. The SDS algorithm is more capable of modeling the complex relationships from features of LR to those of HR, thus improves the quality of reconstruction.As for regression-based exemplar learning methods, some use subspace clustering to improve the prediction performance. However, the reconstruction accuracy cannot be guaranteed, since the clustering analysis is usually conducted only on LR features instead of joint ones. We proposed a coupled subspace projection(CSP) based SISR method to address the problem. Firstly, a projection tree is constructed by maximizing correlation intensity, which is used to partition feature spaces into several correlated linear subspaces. Then for each of the subspaces, a mapping from LR to HR features is learned by anchored point regression. Experiments demonstrate the effectiveness of the proposed CSP method.