Research On Digital Image Completion Methods And Its Application In Image Compression

Digital image completion is a significant research topic in image processing and computer vision. With the development of computer technology, image completion has been widely used in application areas such as image compression, video error concealment, target removal, and so on. Because image completion problem is a typical ill-posed problem, when it comes to deal with it, it not only depends on algorithms and models the method utilized, but also needs to take the general rules of human perception into account. From the perspective of mathematical model in completion problem, this dissertation provides the complexity of the problem and challenges we faced by combining with visual perception. On the foundation of Bayesian Interference, this dissertation also illustrates the ideas, methods and applications related with image completion thoroughly from following aspects.Firstly, a comprehensive introduction of classic methods in image completion is presented, including partial differential equation-based image completion methods and texture synthesis-based image completion methods. By comparing with the two sorts of methods in advantages, disadvantages as well as their corresponding applications, an image completion algorithm based on non-parametric texture synthesis is simulated and improved in this dissertation, which is called Belief Propagation (BP). The algorithm introduces a well-defined objective function under the model of discrete Markov Random Field (MRF), and solves the function by a global optimization scheme.Secondly, in order to improve the completion result of this algorithm, a novel priority scheme is added on the basis of original Priority-BP, which uses the current optimal label candidates of Markov nodes to calculate sparsity of image patches, and then decides structural priority by the value of patch sparsity. Compared with previous patch-based priority scheme of completion, priority scheme in this dissertation can capture structure information in images more accurately.Thirdly, in order to improve the efficiency of completion algorithm, the label candidates in original BP usually have to be pruned. Based on the label pruning criteria of belief value mentioned in Priority-BP, a new criteria of structural similarity is incorporated into label pruning, so that faulty messages which passes among Markov nodes are restricted. Meanwhile, the efficiency problem of BP can also be solved by this new label pruning criteria. Finally, starting with the improved BP-based image completion method, an image compression system at the purpose of eliminate visual redundancy is studied. The system first extracts structural information from original image as assistant information at the encoder side, and then intentionally drops some blocks in original image so as to the incomplete image can be coded and transmitted together with the assistant information. After decoding the incomplete image and assistant information, the image can be reconstructed with the help of assistant information by image completion techniques. Eventually, an image compression technique which can exceed the limitation of Shannon information theory is realized.