Structure Regularization Method For Image Processing

In recent years,the ill-posed problems in image processing have become a hot issue.The most common and effective method to solve these ill-posed problems is the regularization method.Its main idea is to add constraint terms with guiding effect to the objective function,which can make the final solution have the desired properties.Image processing methods based on regularization constraints have been extensively studied,but how to construct ap-propriate regularization constraints is still a challenging problem.In this paper,for the two tasks of image restoration and image classification,we design reasonable and effective reg-ularization models by analyzing the ideal structure of target variables.Specific works are as follows:A weighted-l₁-method-noise regularization model for image deblurring has been proposed.Specifically,the proposed regularization term adopts thel₁penalty to constrain the method noise,and incorporats a gradient based weight.The advantage of the regularization term is that it can remove the noise in the smooth areas while preserving image edges better.Extensive experimental results show that the proposed method can obtain better results than other method noise based regularization methods.A new cooperative game framework for joint image restoration and edge detection is pro-posed.It consists of two objective functions,one aims to detect edges from the unknown real image,the other aims to restore the unknown real image with the supervision of the detected edges.An iterative algorithm for solving the model is given,which makes the two objective functions promote each other in the iterative process,and its convergence is proved with mild conditions.Extensive experimental results show that the proposed method outperforms other related methods in both image restoration and edge detection.A group discriminative least square regression regularization model for image classification is proposed.Specifically,this model adopts the difference ofl_2,1norm andl_2,2norm of the predicted labels of each class to constrain the label transformation matrix,and relaxes the binary label by the existingdragging technique.In addition,an iterative algorithm for solving the model is given.Extensive experiments show that our method outperforms both related methods and some traditional methods.A classification-friendly sparse encoder and classifier learning regularization model is pro-posed.Specifically,the model directly employs the training set as the class-specific synthesis dictionary,and an extra smoothing term is introduced to enforce the representation vectors to be uniform within class.The advantage of the model is that the sparse representation of training samples can be guaranteed to have an exact block diagonal structure.Extensive experiments show that our method outperforms some state-of the-art model-based methods.