Research On Image Super Resolution Reconstruction Algorithms Based On Regression And Sparse Representation

Single image super-resolution(SR),important problem in image processing and computer vision,aims at restoring a high-resolution(HR)image from its low-resolution(LR)input image.Its has wide applications in biometric recognition,medical imaging diagnostics and remote sensing image detection.Based on sparse representation theory and neural networks approximation theory,this dissertation proposes SR reconstruction algorithm based on adaptive sparse representation and self-learning,super-resolution reconstruction algorithm based on adaptive sparse representation and semicoupled dictionary Learning and local linear projection,and image superresolution reconstruction algorithm based on feedforward neural network and classification.The specific content of the work are described as follows:1.The key of regression-based single-image SR problem is to establish a mapping relation between LR and HR image patches for obtaining satisfied reconstructed effect.In view of the strong nonlinear approximation ability of neural network and clustering,we propose a SR reconstruction algorithm based on a new two-dimensional back propagation algorithm(2D-BP).Firstly,we cluster the feature space of LR images into multiple LR feature subspaces and group their corresponding HR feature subspaces.According to the local geometry property learned from the clustering process,we collect numerous neighbor LR and HR feature subsets from the whole feature spaces for each cluster center.Secondly,we construct a twodimensional feedforward neural network to learn the mapping relation from the LR feature space to HR feature space in each subspace.Comparative experiments demonstrate that the method proposed in this paper has better reconstruction effect.2.This paper proposes a novel super-resolution reconstruction algorithm based on adaptive sparse representation and self-learning framework.The fidelity term in the model ensures that the reconstructed image to be consistent with the observation image.While the adaptive sparsity regularization term constraints the reconstructed image with an adaptive sparse representation,which successfully harmonizes the sparse representation and the collaborative representation adaptively via producing a suitable coefficient.To construct a more effective dictionary,the high frequency features from the underlying image patches are extracted,and the dictionary learning and sparse representation are integrated.To this end,the alternating direction method of multipliers(ADMM)and iterative back-projection(IBP)method are applied to solve the optimization.Comparative experiments demonstrate that the proposed method is effective in term of both quality and noise immunity.3.Traditional sparse-representation-based SR reconstruction methods are based on the assumption that the HR and LR image patches have the same sparse representation coefficients over the pair of dictionaries.This paper adopts a semi-coupled dictionary learning framework and proposes an SR reconstruction algorithm based on adaptive sparse representation and semi-coupled dictionary learning(ASC-SCDL).This algorithm framework assumes that there is a linear relation between the sparse coefficients of LR and HR patches under their corresponding dictionaries,which not only relaxes this assumption,but also enhances the mapping relationship between patches.In combination with adaptive sparse representation,the stability of sparse decomposition is guaranteed.During the training phase,K-means classification is performed on training set.Then a strategy based on sparse-domain error reclassification and semi-coupled dictionary alternate learning applied.The semi-coupled dictionary pair and the mapping matrix of each class are trained together.In the testing stage,the semicoupling sparse classifier is used to reconstruction the HR image.Experimental results show that the proposed method has better reconstruction effect.