Bayesian Analysis Of Dynamic Images

Analysis of dynamic images is an integrated application of image- processing approaches, and is the first step of all researches of computer vision.The objective of this paper is to make a deep research and a careful analysis on traditional Bayesian dynamic images analysis methods at first, and then to propose a new more accurate and faster method of dynamic images analysis, which is based on Bayesian statistical theory and multi-resolution technique. Our research work in this dissertation is showed in the following:1) An affine model of image regions' motion is built after the Bayesian image process theory and the Markov Random Field (MRP) theory were studied. The fundamental of this model is assuming that the change of the projection of moving objects on the image plane can be modeled by an affine transformation. Then these parameters of regions' motion will be obtained from this model. When the model is used in the Markov random field, the basic approach of dynamic images analysis, which is named the Adaptive Segmentation Algorithm (ASA), is formed. After that, the initiation of the dynamic images analysis approach is investigated carefully.2) In order to solve motion analysis problems in the Markov random field, some basic Bayesian methods of images segmentation are studied. At first, the principle and the convergence feature of Gibbs sampler are discussed; and then, some fast annealing algorithms, such as Metropolis method, the ICM method and the HCF method, are studied carefully. Among these three methods, the Metropolis method is most closed to the Simulated Annealing algorithm, but it also has a low converging speed like the SA. The ICM method is the fastest one, which is equivalent to the zero temperature annealing procedure, but it will drop into the local minima inevitably in the segmentation task. In order to make the most blurred point converged most lately, the concept of "stability" is introduced into HCF algorithm, and the initiation progress is also embedded into it. Although the HCF algorithm is more accurate than the ICM algorithm, it will cost more computer time. According to the simulation, the problem about local minima in this paper cannot be solved by this method. Because the accuracy of these three algorithms are very closed in our simulation, the ICM method is our most appropriate choice, and will be the basic method in the following segmentation work in this paper.3) For the purpose of the large-size-image segmentation, the improvement of the accuracy and the increase of the computing efficiency, the multi-resolution technique is proposed to be used in the images motion analysis, so as to form a basic structure of the dynamic images analysis frame in this paper. Then the Gaussian images pyramid is built. The accuracy of the image analysis will not be affected under the lower resolution in our analysis work, and the segmentation result in the coarse resolution will be used as an initiation of the analysis in the fine one, which will satisfy the accuracy requirement of the fast annealing algorithms. The most important problem in the multi-resolution method is how to reduce the effect of outliers while the estimation of the motion parameters are translating from the coarse resolution to the fine resolution. To solve this problem, some new estimation algorithms are used in this paper, which are MRSL, IRSL and PSM. Among these algorithms, PSM is the most stable and the most accurate one.4) Some problems existing in the early multi-resolution dynamic images analysis are discussed, and our solution is provided, which results in a new multiscale dynamic images analysis method. In those early methods, the coarse images will be discarded after they are processed. So the scale information in the image pyramid cannot be used completely. In order that the scale information can be used efficiently, and the multi-resolution approach can be more perfect, a new multiscale segmentation scheme is proposed, and the Multiscale Auto-Regressi...