Systems theoretic approach to textured image and video processing .

摘要

Control Theory and Image Processing are all exciting research areas with great power for the applications to wide fields. There are many common methods in system analysis, design and development in these two different fields. The progress of these two area also shows that the techniques developed to solve the problems of one area often find applications to the other one.; The theme of this dissertation is systems theoretical approach to textured image processing and video processing. It is an attempt to setup and solve the textured image processing problem from a viewpoint of control, especially the study of robust identification and robust control theory. The work focus on texture modelling, synthesis, recognition and classification. A novel image modelling and model reduction approach is introduced. It is shown how recently developed robust identification techniques can be applied to find models for textures which are capable of image compression and reconstruction. On the other hand, video inpainting problem is addressed under a framework combining Local Linear Embedding (LLE), Rank Minimization Interpolation (RMI) and Radial Basis Function (RBF) Mapping, leading to a simple, computationally attractive, dynamic video inpainting algorithm. Proceeding along the same lines, rank based approaches are proposed to solve event detection and track matching problem.; The contribution of this dissertation can be viewed both as theoretical and practical: It provides answers to the robust identification of 2-D discrete, quarter causal, shift invariant systems that have a periodic impulse response, which is also of great practical interest in image processing, distributed systems and so on. Moreover, by introducing rank minimization algorithm, a new solution is provided to the problem of video inpainting, which can deal with video inpainting under the conditions of non--periodic target motions, non--stationary backgrounds and moving cameras. Finally, with the idea to detect dynamics changes by parsing it into segments according to the complexity of the model required to explain the observed data, a rank based approach is introduced to solve track stitching and dynamic event detection in a unified way.