Video Retrieval Based On Feature-Fusion

The text-based video retrieval exhausts a great deal of manpower, and the video descriptions are very subjective, therefore the content-based video retrieval (CBVR) has become a hot topic in recent years. The process of CBVR roughly be divided into the following steps:firstly, shot segmentation; secondly, extraction of the key frames; thirdly, extraction of shot features; lastly, video retrieval based the similar shot features.As the video shot segmentation is mainly based on the extracting features, we focus on the commonly used video features, such as color features, main color features, texture features, shape features and video manifold features, and we mainly focus on the Spatial-Temporal manifold features for high-dimension video data using Spatial-Temporal LLE algorithm in this thesis. Experiments show that the repetitive motion video content can be well characterized by the Spatial-Temporal manifold features.The shot is the basic indexing unit, and the shot segmentation is the basic step of the video retrieval. A number of commonly used classic video shot detection algorithms are introduced, including global threshold method, ECR method and so on. We make a careful analysis of the manifold of the shot boundary with the ST-LLE method. We propose a shot segmentation method based on main color descriptor and main color layout descriptor, and the other method based on the histogram changing rate.In this thesis, We mainly research methods of the feature fusion of the video. The retrieval result based on single feature is not very good because the single feature is difficult to provide a comprehensive response to the video's content. We need to find a more effective method to integrate several features to describe the video comprehensively.