The Research Of Image Super-resolution Reconstruction Via Sparse Representation With Mixed Samples

The image super-resolution(SR)is a technique that to reconstruct a visually fine high resolution(HR)image from one or a series of low resolution(LR)images using signal processing techniques.Image SR not only is very useful in some special fields,such as medical imaging,satellite imaging and so on,but also can satify the sustained growth of the demand of HR images.Since it is very signigicative to research the image SR,this thesis focuses on the single image SR.Single image SR is a technique that to reconstruct a HR image from one observed LR image,whi ch is a common and challenging question in image SR field.The main research is as follow: firstly,since it is hard to accurately regularize the local image structures,this theise presents an auto-regression and moving average(ARMA)model-based local structure regularization term.Secondly,due to both the drawbacks of external sample and internal sample,this theise proposes an adaptive sample selection(ASSS)scheme and an adaptive mixed sample ridge regression(AMSRR)model to employ the complementary information included in external and internal samples for better SR results.Finally,due to it is hard to represent the local image features in spatial domain,this theise gives a research on sparse representation in wavelet domain.Due to the extremely ill-posed characteristic of image SR,this theise proposes an ARMA model based local structure regularization term.In training phase,the ARMA models are trained from external samples,and in reconstruction phase,the proposed algorithm adaptively selects the most suitable ARMA models to construct a local structure regularization term.Moreover,this theise considers the nonlocal structure regularization.Both the local and nonlocal regularization terms are unified into the sparse representation-based framework to form a dual regularization term.Compared with the classical sparse representation-based method,the dual regularization term-based method proposed by this theise not only promotes the objective indexes(PSRN and SSIM)by a big margin,but also im proves the visual quality of the reconstructed HR images.The example-based learning SR algorithms can be mainly divided into the external sample learning methods and the internal sample learning methods.The external sample learning methods usually suffer from the influence of the similarity between the images in training set and the input LR image.While the internal samples usually suffer from the unitary samples,as well as the learned mapping relationship s,which can't represent various image structures.Due to both the drawbacks of external and internal samples,this theise proposes an ASSS scheme and an AMSRR model,as well as the corresponding optimization algorithm to employ the complementary information.Extensive experimental results demonstrate the effectiveness of the proposed algorithm via comparing with the state-of-the-art methods.Due to it is hard to represent the local image features in spatial domain,this theise gives a research on spare representation in wavelet domain.Firstly,the ima ges included in the training set are directly transformed via Discrete Wavelet Transform(DWT).Four wavelet sub-band HR and LR dictionary pairs are trained via K-Singular Value Decomposition(K-SVD).The trained HR and LR dictionary pairs are used for the HR image reconstruction in wavelet domain.Compared with some spatial domain methods,the proposed wavelet domain method is effective in computing and omit the feature extraction operator.Experimental results verify the effectiveness of the proposed wavelet domain method,which outperforms the comparing spatial domain methods both objectively and subjectively.