Research And Application For Image/Video Segmentation Technology

Segmenting image/video into meaningful regions and then summarizing them are both important and key technologies of automatic catalogorization for digital library. Image/video segmentation is one of the most active research topics in computer vision. It is a crucial step from image process to image analysis and it is also the basis of vision tasks such as image object s'extraction, measurement and recognition.In this thesis, a brief overview of current main image segmentation technique is presented, and then three basic pixel-based image segmentation algorithms are discussed. They are:First, Expectation-Maximization algorithm based image segmentation, this method assumes that color space of image pixels can be approximated to mixture Gaussian distribution, and parameters estimation of the mixture Gaussian that the number of Gaussian components is given can be achieved by typical EM algorithm; Second, Markov Random Field algorithm based image segmentation, image can be modeled as Markov Random Field, then it is transferred to the problem of how to make Gibbs energy function smallest; Third, region growing based image segmentation, defining certain similarities measurement and gathering pixels who have similarities on certain rules.Further, base on analysis and discussion of the three algorithms, an automatic image segmentation technique is proposed:EM-MRF automatic image segmentation algorithm for gray-level images. Algorithms using Matlab/C mixed programming technique, Matlab has become an important tool for researchers because of its easiness and graphics processing capability. However, calculation speed limits their application to large calculation such as image segmentation. Therefore C language is used to complete parts of iteration operation in image segmentation algorithms, base on it use Matlab to achieve all algorithms researched. A large number of experiment results validate strongpoint and shortcomings of the three algorithms aforementioned. EM-MRF proposed in this thesis has been tested, experiment results show that this algorithm can segment images automatically and effectively and it can outperform EM or MRF.Finally, we propose a Gaussian mixture model for the video segmentation, based on our image segmentation techniques. Different from previous video segmentation techniques which segment video in frame level, the proposed method performs segmentation in both spatial and temporal domain. Experimental results validate the effectiveness and performance of the proposed methods.