| Dynamic textures (DT) are sequences of images showing spatial regularity, temporal continuity and infinity and comply with certain statistic characteristics. Many phenomena like flowing water, wind-shaken trees and flags etc. can be classified as examples of dynamic textures. Research of dynamic texture mainly consists of analysis, recognition, segmentation and synthesis of dynamic textures. Among them dynamic texture recognition has the most magnificent importance both theoretically and practically, which is also what we are concentrated on in this paper. Our purpose is to compare and find out some method that can describe and classify dynamic texture with as higher recognition rate and less computational cost as possible.Creating a model for dynamic texture is the first step in both dynamic texture recognition and synthesis. A variety of methods of modeling are employed in different algorithms. In Chapter one of this paper, we first research on Linear Dynamic System model (LDS model) which is the most representative and of widest application. The LDS model used a one-order auto-regressive equation to describe the dynamics of the DT and use linear PCA algorithm to reduce the dimensionality of the high-dimensional image. The LDS model is easy thus accept wide application in dynamic texture recognition, synthesis and segmentation. However, LDS model cannot capture complex appearance changes because its linear dimensionality reduction scheme is too simple and has poor non-linear features. Therefore in this paper, we introduce a novel modeling method—NLDS(Non-Linear Dynamic System) model to improve the traditional PCA based LDS model. NLDS model has been proved to be effective in recognition of human faces, handwriting and so on. We use this method for a more precise description of dynamic textures and simulation result of MATLAB also reveals a better performance of the new model.In chapter 2, we researched on Martin distance algorithm, which is a distance-based method used for differaciate AR models. Compared with the other two distance based algorithm, say KL distance and geodestic distance, Martin distance can achive higher recognition rate with relatively lower computational cost. Therefore, in this chapter, we are focused on Martin distance algorithm. Furthermore, we combined it with KPCA and deduce the NLDS based Martin distance recognition algorithm. Finally, experiment is also carried out for evaluating the performance of this method. In Chapter 3, we research on a novel, non-distance-based recognition method called impulse response based method. Compared with those method discussed in Chapter 2, this method requires much less computational cost and has better performance in recognition rate. It differentiates dynamic textures with their own impulse response, which is a commonly used method in system identification. However, this method was first proposed based on LDS model which is not that precise anyway, so we combine our method of NLDS model and impulse response together and realize a trade-off in computational complexity and recognition rate. Simulation result shows that the proposed algorithm is more accurate in modeling dynamic textures and the recognition rate has an increase of about 20% compared with the original method. Finally, in order to overcome the flaw that the impulse response algorithm neglects the appearance of dynamic textures, we introduced the LBP algorithm and figure out corresponding solution and support it by simulation. |
