Research On Methods And Applications Of Digitial Image Sparse Representation

The representation of a digital image is a fundamental problem in image processing.Image processing and its applications are based on the efficient representation of image content.Sparse representation is a novel theory of image representation theory which can represent an image concisely.Sparse representation has attracted a great deal of attentions and has an important impact on signal processing and analysis.At present,it has been widely applied to many fields such as image processing,pattern recognization,automatic measurement and control,etc.This thesis addresses the issues of sparse representation and its application to image processing.The main research works and innovations in this thesis are summarized as follows:(1)This thesis proposes a denoising algorithm based on the size of the dictionary and sparseland model-based image geometry,combined with adaptation.Firstly,contrast four kinds of image geometric features dividing method,we find the right division method.Gaussian smooth gradient map may accurately distinguish the detail,texture and smooth area.Then DCT dictionary is used as a sparse dictionary,experiments show that with the change of adaptive block size,the peak signal to noise ratio and the "visual effect" are different.The denoising inconsistencies caused by the use of DCT dictionary lead us to use K-SVD dictionary as sparse dictionary.For K-SVD algorithm,experiments show that the way to initialize dictionaries is very important.Experimental results show that,it is better to use a randomly selected image block as the initial results of the initial dictionary rather than use DCT.(2)Based on the theory of FPC,combined the fixed point formulation and preconditioned conjugate gradient method,this thesis puts forward a new recovery algorithm and proves its convergence.Through the sparse signal recovery,and contrast some existing algorithms,the proposed method may get less norm errors and faster speed.The algorithm is applied to image deblurring,image separation and image restoration problems,and achieves good results.(3)Base on Compressive Sampling Pursuit(Co Sa MP)algorithm,this thesis proposes an algorithm which select a few atoms of overcomplete dictionary that produces discrimination in representations of training images.The representations can be used as low dimensional features of images to classify images.Firstly,the thesis proposes a new combination of sparsity and discrimination measure of the objective function.In order to verify the effectiveness of the objective function,combined Co Sa MP algorithm,the experiments show discrimination Co Sa MP less than rebuilding in the case of equal accuracy needed atom.Using two different sets of experiments to test the proposed algorithm performance and robustness.To verify the robustness of the algorithm,the algorithm is noiseless operation under conditions different levels of noise and occlusion.In addition,experiments show that the algorithm can handle both high intra-class variability image data and low intra-class variability image data.The proposed method has higher accuracy than LDA using smaller number of features.(4)The thesis proposes a feature selection algorithm that combination of discrimination metric,compressed measurements,for classification purposes.The proposed algorithm removes the original compressed sensing redundant feature for improving classification accuracy and decreases the sparsity.Noise-free TU Darmstadt database is used for testing the proposed algorithms,the experimental results of using a subset of the feature set are similar or better than the results of using all group features.(5)Based on the iterative hard threshold(IHT)algorithm,the thesis proposes a new joint sparse representation model reconstruction algorithm.The proposed algorithm has a low computational complexity.To the proposed algorithm,the signals can be sensed by different sensing matrix.The calculation time of the proposed algorithm is less than that of Duarte's algorithm while they have a similar performance.With the increase in the number of iterations,the proposed algorithm is robust to noise ratio and reconstruction.