Facial Image Compression Based On Dictionary Learning

Digital image compression is related to many theory and technology, such as signal expression, pattern classification and quality evaluation. It is not only a basic problem of digital signal processing but also needed to be settled immediately. This thesis focus on the digital image compression related topics, meaning over-complete representation, the combination of complete and over-complete representation and selecting compression algorithms based on the blocks'attribution.The over-complete representation of digital signal is developed and became familiar in recent years. The traditional complete representation based algorithms cannot fit some situation since they do not consider the internal attribution of the signal data. The over-complete representation progress from the pre-determined dictionary to the learned dictionary and represent the signal data better. This thesis analyzes different match pursuit and dictionary learning algorithms, designs a coding method, based on which the following proposed algorithm's advantage can be judged. Comparing the revised algorithm with the original ones and setting a standard for the further research.Then we compare the advantages of complete representation and over-complete representation in the signal compression's field. We explore the possibility that combining these two structures. We propose and prove three theories and develop a new digital image compression method. The method takes into accounting the image's sparsity under complete representation and over-complete representation, brings into effect a new way to code the coefficients, tries and proves the effect while coding with the dictionary and its derivative. The experiment validates the proposed algorithm.Selecting compression algorithms based on the blocks'attribution is given birth based on this observation that different representation and compression methods all act well in some particular situation. The PSNR and processing time are the two requirements we need to make balance of. On the other hand, once we use an algorithm to deal with some images'compression just because it fits most of the data's attribution. But the complicacy of the real images means this algorithm cannot advance in all regions of the images. Using different based on the attribution may be a solution. We do research on this possibility and the suitable situation. We also propose a real compression system's construction. As a conclusion, we review the thesis as a whole briefly. We also draw the expectation of the following research, meaning collecting a much more big and complex database, using the latest result in compressive sensitive to design new dictionary learning method, selecting the right dictionary quickly, etc. Then we may find a new way to boost the PSNR with the same bpp and reduce the processing time further more.