Deconvolutional Networks For Image Representation And Restoration

Image representation is a basic and key problem of image processing and analysis. Also, it is one of the most important and difficult tasks in computer vision. Normally, the familiar approaches for image representation by transformation with the data-level always ignore image contents, and it can't match demands with the mid-level and high-level tasks. Extendable image representation based on image features can solve this problem commendably, but it can't be applied to image reconstruction which is a low-level task in computer vision.Deconvolutional Networks model is researched in this thesis, which can get mid-level and low-level image features by training and learning with the hierarchical structure. Meanwhile, feature representation and reconstruction can be learned with the overcomplete Deconvolutional Networks. Therefore, the approach studied in this paper can solve problems with image representation approaches by transformation and based on features. For details, the thesis unfolds the work as follows.Firstly, this thesis introduces and studies the Deconvolutional Networks model based L₁ -Regularization. Sparse image representation with mid-level and low-level features is studied in detail. Also, the thesis analyzes the traits of reconstructed image for future study. The thesis designs an image representation and restoration algorithm for remote sense images with the Deconvolutional Networks model based on L₁ -regularization which is firstly introduced to remote sense domain. Extensive experimental results illustrate the good performance with the model which is applied for feature extraction, sparse representation and remote sense image restoration,Secondly, the thesis analyzes the sparse extent with the different regularization for the Deconvolutional Networks model and introduces a Deconvolutional Networks model based on L_1/2.By theoretical analysis and experiment results, the Deconvolutional Networks model based on L_1/2 can extract more image features and keep more similarity which can be applied to image restoration than the Deconvolutional Networks model based on L₁ .Lastly, the thesis concludes the strong points and drawbacks of the researched approaches and puts forward directions for the next research. It should be paid more attention that the theoretical excellence and application potentials with the researched approaches are performed by image restoration tasks, but the approaches have an unified framework which can be applied to other image processing and analysis tasks.