Segmentation Algorithm Based On The Mpeg-4 Video

Audio－video object－based coding and description are the district features of The Mpeg-4 standard．A video sequence must be segmented into a set of video object. Algorithms of video segmentation that are good adaptive character and real time are crucial to the final MPEG－4 coding products．There exist the contributions in this dissertation as follows．Obtaining the edge of video objects is crucial to the video segmentation, author researches the features of human being visual, especially giving a description of Mach band that is usually used in the image processing, and builds a model of mathematics, researching the Mach band of color image. The article plots gray and color difference curves based on Mach band images. Author introduces an algorithm of automatic threshold choice based on the curves in the wavelet edge detection. Gray and color images are tested, and there is a good result. There is a problem in the wavelet edge detection that people cannot solve so far. People do not know to how to choice wavelet scale to obtain a good edge. Author suggests a new method of edge classifying and researches the deviation of edge points by the difference of scale and edge styles in the wavelet edge detection. As a result, an algorithm of using bigger scale is suggested in the wavelet edge detection. Experimental results demonstrate the performance of our algorithm. Author has studied many kinds of video segmentation algorithms, and suggests an algorithm of semiautomatic video segmentation based on characteristics of visual to improve other algorithms. Users plot a coarse outline of video objects in the graphic user interface (GUI) using the mouse at the first step, then fill the outline to obtain a binary model, using seed growing and wavelet edge correct the outline. In tracking video objects, we obtain an initial segmentation uses motion information and the model of previous frame, and correct by the information of space. Finally, we obtain an accurate segmentation. Experimental results demonstrate the performance of our algorithm.