Image Super-resolution Reconstruction Based On Group Sparse Representation

Super-resolution (SR) image reconstruction is currently a very active area ofresearch, and SR technology gets a fast development, as this algorithm becomes oneof the main forces to breakthrough the limitations of hardware, so that reconstructionimage can be correctly recovered from the down-sampled single.There are many methods of image processing, mainly work are focused on twoaspects: image denoising and image reconstruction.In terms of image denoising, according to the type of noise, the mainstreammethods are median filtering in the spatial domain denoising algorithm, K-SVDdenoising algorithm, wavelet threshold denoising algorithm, PDE denoisingalgorithm and variational (TV) image denoising algorithm. The K-SVD algorithmcan improve the denoising effect by sparse coding atoms instead of the traditionalmethod of sparse coding dictionary.In terms of image reconstruction, reconstruction algorithms are mainly dividedinto three categories: the method based on interpolation, the method based onreconstruction and the method based on learning. Currently, the reconstructionmethod based on learning becomes a hot research. It has an excellently effect toobtain the hiding details of texture from the low resolution image by learning thesimilarity of high-frequency information. Compared with by artificially increasingthe constraint entry, this method has a stronger adaptive capacity.However, the reconstruction effects are affected, as the algorithms mentionedabove ignore that the structure of the dictionary has an effect on the reconstructedimage. Considering that traditional algorithms does not take into account thestructure between adjacent atoms, and ignores mutual constraint between atoms. Tofurther enhance the image processing results in these two areas, the mathematicsmodel of the dictionary learning based on the group sparsity will be constructed.This paper will use group sparsity method to deal with the problem of imagereconstruction and image denoising. The key point of this method is dividing the sparse coefficients into groups and adjusts the correlation among the elements bycontrolling the size of the groups, which reflects the texture characteristics of theimage. The main contents are as follows:(1)Firstly, the paper will introduce the super-resolution image reconstructiontheory and theoretical models sparse representation.(2)Image denoising based on group sparse representation over learneddictionary. As traditional sparse representation algorithms not consider the structureof image, and ignore the correlation between adjacent atoms, this chapter uses theGOMP algorithm to deal with the problem of image denoising. This new method hasa better effect to prevent the "block effect” of recovery image, its texture are moresharpness.(3)Infrared image super resolution via locality-constrained group sparserepresentation. Locality must lead to sparsity, but not necessary vice versa.Considering group sparsity method can not ensure the sparsity of sparse coefficient,we add locality constraints to remedy this defect.(4)Super resolution based on locality-constrained dynamic group sparsity ininfrared image. Considering the traditional group sparse coding usually required toprovide a priori knowledge: Group structure (such as group size/numbers/location). In practical use, we are often unable to predict a group structure of sparsesignals, but only can rule the group size, numbers or location by artifact. Thedrawback is not flexible enough to divide the groups according to the given image orthe signal and has a high complexity. Dynamic group sparse coding can dynamicallyhandle sparse representation coefficients without knowing groups structure. It bringsa great convenience to sparse processing the signal.