Image De-blurring Based On Adaptive Sparse Model

Image information closely related with people's lives,but the influence of various factors in the process of acquiring the image often resulted in artifacts,such as blurring and noise,seriously affect the subsequent processing.Image deblur technique can improve the image quality effectively and achieve the recovered cleared image.In the front part,we introduced the essential principle concept and model,including image blurring process model,image de-blur model,the ill-posed of image recovery problem and how to resolve it.Then we emphasize on the importance of image priors which can significantly enhance the recovered quality.We give the methods to evaluate image recovery quality,including subjective evaluation and objective evaluation.The main contents of this thesis can be summed up in two aspects:1)According to the characteristic of image block stableness,we present the adaptive deblur model in different image texture,based on variable split framework,and demonstrate the variable split framework also applies to our adaptive improvement.Assumed that image gradient meets the parametric generalized Gaussian distribution and build the adaptive model under the MAP framework.An effective GGD parameter estimation method is given,we prove that the global adaptive model can't get the best result of deblurring,so present a semi-adaptive deblur model.A brief process is described as follow:first we get the intermediate result by pre-processing the blurring image,present an effective and simple method to partition the image into the texture areas and smooth areas.In texture region we use a globally convergent algorithm(GCM)to estimate the parameters of GGD,but set a fixed parameter value in smooth region based on experience.By comparison with several other most advanced algorithms,it shows that our adaptive algorithm can achieve better visual quality and higher signal to noise ratio.2)An adaptive analysis sparse model based on image texture is presented by taking advantage of natural image nonlocal self-similarity.The model first extracts all images blocks from the current image and uses k-means clustering algorithm to divide the data into M clusters.In each cluster,we use the GOAL method to learn M dictionaries,each corresponding to a particular texture.In the calculation process,each image block selects its nearest match dictionary automatically according to the Euclidean distance between cluster center and image block,so that it can achieve a better sparse representation.In order to enhance the quality of recovery,we introduced a new non-local spectral priori(NSP)model for different texture.The model learns different NSP model from natural images according to various image texture.Combined with the analysis sparse prior,we propose an adaptive mixed prior model,and give the specific algorithm to solve it.The simulation and experiment results show that the proposed deblur model can improve the adaptability of local texture and achieve better image restoration quality.