Dictionary Learning Method In Sparse Representation And Its Applications From Remote Sensing Image

Sparse representation, which aims to capture the essential features of signals, is a fundamental and critical problem in the areas of signal processing, especially in the field of high-resolution remote sensing image processing. When the remote sensing data is large and complex, sparse representation can remove the redundant information, and reveal the innate character, which is important for the subsequent processing. It has recently received a lot of attentions in a wide range of applications including remote sensing image classification, coding and fusion, etc. However, due to the complex of data and due to various applications, it is hard to design a universal over-complete dictionary for solving sparse representation tasks. Therefore, designing an efficient over-complete dictionary applied to high-resolution remote sensing image sparse representation is a very interesting and challenging research topic.Based on the current trends of over complete dictionary, this dissertation aims to research reconstruction dictionary and discriminative dictionary for high-resolution remote sensing image classification and coding. Based on the sparse representation theory and dictionary learning methods, we have designed two types of dictionaries. As for reconstruction dictionary, we focused on the sparse performance, and the rebuilt quality; as for discriminative dictionary, we focused on the divisibility of the coefficient matrix. The main work of this dissertation is summarized as follows:Firstly, based on the theory of reconstruction dictionary, a layer segmentation based dictionary learning model is proposed. In order to reduce the texture distortion and geometric distortion of image, the original image is divided into edge layer and residual layer. The edge layer is represented by edge-related atoms, which generated by some generating function and geometric operations. The residual layer is represented by a learnt dictionary. Furthermore, in order to improve the algorithm’s complexity, the improved SVD and improved OMP algorithm are used in dictionary learning. The reconstruction dictionary in the dissertation does good performance of reducing rebuilt image distortion.Secondly, based on the theory of discriminative dictionary, a layer segmentation based discriminative dictionary learning model is proposed for remote sensing image classification. The model added term of coefficient divisibility constraint into the object function of reconstruction dictionary learning model. In order to optimize the object function, the forms of the constraint terms are kept uniform. And traditional K-SVD is used to solve the optimization problem. The discriminative dictionary in the dissertation is compact and the experimental results have confirmed that its coefficients are more divisible.Then, based on the discriminative dictionary in this dissertation, a sparse representation based classification model (SESRC) is proposed. A classification error constraint term is added into the object function of discriminative dictionary learning model. Also, as in the discriminative dictionary learning model, the forms of the constraint terms are kept uniform. And traditional K-SVD is used to solve the optimization problem. SESRC learned the dictionary and classifier parameters as the same time. It is not a two-step as in many other algorithms. The experiment has shown it had good results in classification Accuracy and computation time.Finally, to solve the high resolution remote sensing image coding problem, a classification based coding scheme is proposed. This scheme is a two-step program. Firstly, the images are classified into several ground feature species. Then different types of image are coded separately. This coding scheme can achieve a fixed bitrate compression with the given quality requirement. In addition, we propose a remote sensing image compression quality prediction model for onboard coding.