Single Image Super-resolution Based On Joint Local And Non-local Priors

As one of the basic tasks in image processing,super-resolution(SR)is widely used in many applications,such as medical imaging,remote sensing and entertainment.This technique is able to enhance the resolution of the image without upgrading the existing acquisition systems.Therefore,it is meaningful to study the problem of image SR.The traditional SR methods utilize interpolation to super-resolve the low resolution(SR)images.Due to low complexity,the interpolation-based methods are widely used.However,the quality of the SR result is relatively low.Compared with the interpolation-based methods,the reconstruction-based methods,which achieve better results,receive much attention.This kind of methods relies on the usage of prior information.However,most of the reconstruction-based methods do not effectively make use of the local and non-local information,and therefore leads to a poor performance.To solve the above problem,this thesis proposes two SR methods by jointly exploiting local and non-local prior information:(1)A SR algorithm based on Non-local Total Variation(NLTV)and Steering Kernel Regression(SKR)models is proposed.Firstly,the original kernel regression is utilized to calculate the initial estimation of the kernel.Secondly,SKR is used to modify the estimated kernel,so as to accurately sense the contents in local areas.Then the adaptive-shape block matching is carried out according to the obtained kernel.Finally,the coefficient of the NLTV model is computed based on the adaptive-shape blocks.It is demonstrated through experiments that,compared with the existing methods,the proposed one achieves better results in terms of both objective and subjective qualities.(2)In the process of constructing the ASNLTV model,a non-local Means(NLM)model with low complexity and adaptive shape is also proposed.In this paper,a joint local and non-local priori image super-resolution algorithm is constructed by combining SKR and NLM algorithm.On one hand,similar to our first algorithm,SKR model is used to improve the normal NLM model,leading to the Adaptive-shape Non-local Means(ASNLM)model.On the other hand,the ASNLM model can also be applied to improve the performance of the normal SKR model,leading to our improved non-local steering kernel regression(I-NLSKR).The final HR result is achieved by solving the optimization problem which contains both ASNLM prior and I-NLSKR prior.