Research On The Key Technology Of Sparse Representation Based Image Recovery And Enhancement

With the rapid development of information and multimedia technology,visual data are increasing at an explosive rate,which has significantly influenced people's social life in various perspectives.Dealing with the big visual data,how to discover the valued information and make it play an important role in city security,traffic surveillance,and medical treatment is a very serious topic.However,in real applications,visual data will inevitably be corrupted and degraded,this degradation would affect the usage of the data,so it is a very important topic to seek an effect way to enhance the quality of visual data.This dissertation is focusing on two aspects of the image recovery and enhancement,the first one is to do the image recovery for those images corrupted during the acquirement and transmission,the second work is to remove the rainy effect of the rainy image.In this dissertation,both of these two problems are based on sparse representation for image normalization.For the first problem,we characterized the sparsity structure of the image under the wavelet tight frame;for the second problem,we use adapted dictionary learning method as the sparse system,and try to discover the different structures between the clear image layer and rain layer to separate the rain from the rainy image.The main contributions of this dissertation are listed as follows:1.A novel idea about the sparse representation has been proposed,nowadays,most sparse representation based normalization models seek the position of non-zeros of the image coefficients under some transform,but in our dissertation,we seek the position of zeros and try to build a complete system to recover the image.2.A new image normalization model has been proposed.Most of the currently normalization models with the local priors,they used directly the L₁ norm for the penalty to give the sparsity constraint without the leverage of the connection of the sparsity structure.This dissertation carried the connection between the pixels on the edge to the transform domain,and with this connectivity,more accurate positions of zeros can be detected and more clear image can be recovered.3.A new composite model has been proposed.Most of the currently rain removing works used the linear additive model as the rain composite model,in this dissertation,a new screen blend model has been proposed.This model can adjust the effect between the background and the rain layer and is more real to rendering the rainy image.4.A new image de-raining approach has been proposed,unlike the most used two-step methods,we don't need to detect or do the in-painting.The proposed single image rain removing via discriminative sparse coding has skipped these two steps and considering the rain removing as the signal separation problem.During the separation,mutual exclusivity has been used to characterize the different structure of the clear image and rain layer and the rain was separated automatically and effectively.