Research On Techniques Of Content-based Video Segmentation

With the rapid development of multimedia and network technology, people can get large amount of video information more and more easily. How to organize, manage and retrieve the information has a direct influence on the efficient use of video data. Content-based video retrieval (CBVR), a new research field, is just put forward to resolve this problem and is becoming to the hotspot in multimedia domain. Video segmentation is a fundamental and important step in CBVR.Video segmentation is composed of shot boundary detection and scene segmentation. Based on analysis of the features of abrupt shots, a new novel model called sliding-window difference is proposed, and the method on how to compute shot boundary coefficient is derived from this model. A novel feature selection mechanism is adopted via using shot boundary coefficient combined with HSV divisional histogram. The self organizing map neural network is chosen as the classifier, whose input is the feature and shot boundary coefficient. The set of cut shots can be obtained using this classifier. The joint histogram of angle combined with modulus is selected to detect gradual shots. The joint histogram is computed from the gradient of the low frequency, which can be achieved through a two-layer wavelet transform. The feature is the absolute difference of this histogram. In order to obtain more accurate detection results, an additional morphological opening operation is applied. The key frame extracting algorithm is based on the group of shot boundary. Analyzing color structure histogram of each key frame with the method of fuzzy clustering, scene segmentation can be realized.The experimental results show that the proposed algorithm for cut shot boundary detection is very robust and has high recall and precision for any videos, and the algorithm for gradual shot boundary detection has an good effect on not only some specific gradual types, but also for all of gradual types. Although the performance of the approach for scene segmentation is good, further research for improvement is needed because a threshold which is not automated must be adopted in fuzzy clustering.