| Summary: | 碩士 === 中原大學 === 電子工程研究所 === 91 === Abstract
As the development of MPEG-4 video compression standard, the concept of video object becomes more important. MPEG-4 provides standardized ways to encode video objects, and the scene description, which indicates how the objects are organized in a scene. One of the most important innovations that MPEG-4 brings is the capability of manipulating the individual objects in an image sequence. However, in MPEG-4 the decomposition or spatial-temporal segmentation of a scene into objects is not standardized. Therefore, many object-based segmentation algorithms have been proposed in the literature recently.
These segmentation algorithms use different sets of techniques and result in different performance. Although many approaches try to evaluate image segmentation quantitatively, a universal algorithm for segmenting images and a general criterion for the evaluation of segmentation results do not exist, and most techniques are tailored to particular applications. In this thesis, we propose an entropy-based cost function, which is the weighting sum of each video object’s shape entropy, motion entropy, and texture entropy in a visual sequence to quantitatively predict the coding efficiency of MPEG-4 for a particular segmentation result.
In the experiment, we comparatively evaluate several different segmentation approaches (Watershed and Mean-shift methods, for example) for object-based video segmentation. Experiment results show that the expected results of using the cost function coincide with actual coding results. Therefore, we have verified that by this cost function, we can easily predict which segmentation algorithm can have better coding performance for MPEG-4 video compression.
keywords: MPEG-4, Image Segmentation, Image Compression, Video Object, Entropy, Evaluation Criterion
|
|---|
