Research On Content-based Video Structure Analysis

With the development of multimedia technology and computer network, the content-based video retrieval system becomes more and more important to organize, index and retrieve the massive video information in many application domains. Content-based video structure analysis is a crucial issue in content-based video retrieval and has attracted attention of many researchers in recent years.Video structure analysis aims to map continuous media into discrete and manageable units through segmentation, and therefore serves as a fundamental step to following content representation, indexing, browsing and retrieval. In this dissertation, research efforts are concentrated on the above topic and its related technique, which involves feature extraction, shot boundary detection, semantic scene detection and video abstraction. Several novel and effective algorithms are proposed and validated. For feature extraction, the author proposes a qualitative camera motion classification approach. As Support Vector Machines (SVM) has a very good learning capacity with limited sample set without incorporating problem domain knowledge, in the first step, SVM is employed to classify camera motion operations into two classes: translation and non-translation operations. Then, rotation and zoom operations are distinguished using motion vectors' location and direction. The direction of translation operation is also identified.For shot boundary detection, a hierarchical shot boundary detection approach is proposed. First, singularity detection with wavelet is employed to identify the potential shot boundaries. Then the intensity of motion intensity and block-based edge direction histogram are utilized to distinguish shot boundaries from the false positives caused by rapid motion and complex flashlight scenes. At last, gradual transition is further classified into dissolve, wipe and fade in/fade out.For semantic scene detection, an effective approach based on film grammar is proposed, which gets more reasonable segmentation results, and reduces over segmentation effectively. In this chapter, two important semantic scenes, namely, dialog scene and action scene are identified. At last, the detected action scenes are...