Super Resolution Reconstruction Algorithm Based On Sparse Multi-feature Dictionary With Its HR-LR Synchronous Relevance And Non-local Self-similarity Prediction

A super resolution(SR) reconstruction method improves image quality at lower cost for signals degraded by sensing elements, it provides rich details for applications such as processing of aeronautics and space images, detection of suburb and military targets, surveillance of urban transportation and public security, besides biomedical imaging. Sparse coding based SR(SCSR) is a hot topic on SR researches for its atom descriptors and its dictionary learning, formatting a priori information and increasing sparsity for further representation, respectively. By equal-coefficient constraints to a high resolution(HR) patch and its low resolution(LR) version, SCSR could search the optimal atoms in one or more dictionaries and estimate the HR patch through liner combination of atoms. Our work is focused on:A dictionary trained by flexible LBP descriptors and a multi-feature dictionary tree are proposed to accelerate the atom matching. Firstly, two dictionaries, the edge oriented and the texture based, are created correspondingly using gradient and LBP descriptors to deal with weak structural discrimination in single dictionary case. Secondly, tree structures are introduced to represent the hierarchical clustering of atoms, adaptively choosing dictionary for different structures and leading to a rapid atom searching. In addition, bilateral total variation(BTV) regularization is employed for fidelity of edges and textures.A learning method with combination of dictionary and mapping relation is suggested to increase accuracy of estimating HR sparse representation, which is low because of equal constraint between HR and LR representation. Initially, training HR/LR dictionary and mapping relation of LR-HR representation synchronously, to guarantee relevance of HR and LR representation version from the same patch. Moreover, sparse coefficient reuse is used to modify updating stage of dictionary, for reducing training complexity effect caused by adding mapping relation item. Finally, HR result is obtained by combining LR representation and LR-HR mapping relation, in which the accuracy of estimating HR patch is improved.A joint similarity model and feedforward information based soft-decision scheme is provided to consider fuzzy matching in hard-decision reconstruction assuming unitary relation of HR and LR patch. Foremost, local similarity model is used to reconstruct smooth patch, for recovering simple structure fast and effectively. Furthermore, texture or edge initial HR representations are product of their LR representation and LR-HR mapping relation, and used to build non-local similarity model; weighting similar representations to get target solution, to include more accurate features due to interaction of inside and outside information. Lastly, reconstructed signals are regard as feedforward information for SR model of later patches, to take full advantage of image's inner information.