Image Processing Method Based On Dictionary Learning Research

In real life,we get the most direct visual form of information is the image.Sparse representation is very important for the classification of images,among them,the choice of the dictionary is still a problem worthy to be discussed.In the field of image processing and pattern recognition,the learned dictionary can be applied to the sparse representation in a more accurate and simple form.The purpose of sparse representation dictionary learning is to learn a compact representation of the original training sample data and has stronger recognition level of the dictionary.In recent years we began to refer to the data and information to learn their own complete dictionary,this dictionary by learning to be dictionary atomic number will be more,form and state more colorful,can be better with information and corresponding to the image itself construction,sparse representation more fully.In recent years,through analysis and analysis,we can see that the images obtained by dictionary learning have excellent performance in terms of dryness,trimming and super-resolution.Core contents and directions of this paper can be summarized in the following three aspects:Aiming at the problem that the sparse coding of similar samples is not considered to be related in dictionary learning,an algorithm of relevance constraint dictionary learning based on self-regulated learning mechanism is proposed.The structure of this algorithm is built on its own debug learning system formed,The first step is to use this system to find the most simple sample,and then each selected samples are metaface dictionary learning algorithm as the beginning of the matrix,at the same time the same category sample correlation between adjacent two sparse coding become a new restriction in the project is introduced into the original sparse representation,finally,extended experiments were carried out on the Extended Yale B and AR face datasets,and then comparedwith the other classical dictionary algorithms,the experimental results show that the recognition rate is higher and the robustness is higher.Aiming at the super-resolution reconstruction needs to build a large block library to learn a complete dictionary,and usually very cumbersome over-complete dictionary in the sparse decomposition has hidden uncertainties,easy to produce visual artifacts this phenomenon,an image super-resolution reconstruction algorithm based on weighted classification of block classification is proposed,the algorithm firstly divides the image blocks into the detail blocks and the contour blocks by calculating the variance among the image blocks,then,the k-means is used to further cluster the detail blocks,and the dictionary is expanded for all subclasses to improve the efficiency of the dictionary.At the same time in the image re-establishment process,according to each image block contains details of how much to give different sparse constraint weight,finally,according to non-local similarity approach to expand the follow-up treatment,thereby enhancing the reconstruction effect.The experimental results show that the proposed algorithm preserves the details of the reconstructed image from the viewpoint of subjectivity and objectivity,reduces the deviation between reconstructed image and real image,and improves the reconstruction quality.Aiming at the problem of dictionary learning algorithm discriminant ability is weak,it proposes a kernel dictionary learning algorithm of discriminant based of Fisherface.It combined the feature extraction and dictionary learning algorithm,the first will get the projection matrix combined with PCA and FLD,and the projection matrix as the initial matrix of kernel dictionary learning algorithm;And in order to make the dictionary more discrimination,provide a similarity constraint error term and add in the objective function of kernel dictionary learning,the final classification by the minimum mean square error.In ORL and AR,Extended Yale B three faces library for experiments,the experimental results show that the recognition rate of the proposed algorithm is better than that of Gabor feature sparse representation classification algorithm(GSRC),K singular value decomposition algorithm(K-SVD),and other several classical algorithms.