Content-Based Video Retrieval For Violent Scenes

With rapid development of the computer, network and multimedia technology, people are exposed to more and more video information. How to retrieve useful information needed from the huge video library that contains large amounts of information is a difficult problem which needs to be solved urgently. Currently, content-based video retrieval (CBVR) is becoming one of hot research topics of the image processing research field. It mainly takes advantage of image processing and artificial intelligence technologies to analyze and process the contents of the video. Content-based video retrieval for violent scenes is widely applied in video classification, video cut and public security. The video retrieval and extraction of violence scenes that contains the content of explosions and shootings are mainly researched in this paper.Firstly, the methods of the video segmentation are researched. The characteristics of video data and video structure are analysized. After a study of typical video segmentation algorithms, the dual-threshold algorithm and the histogram-based fuzzy C-means clustering algorithm (HBFCM) are implemented. Then an improved algorithm for HBFCM is designed. Employing the improved algorithm, melting images are eliminated, and the key frames which contain low entropy are reduced substantially. The experimental results show that the improved algorithm is better for video segmentation.Secondly, the content-based image retrieval technologies are researched. The low-level features of images are analysized, such as color, texture and shape. The image retrieval algorithm based on color cumulative histogram is implemented, and is used to retrieve images included in explosion scenes. The image retrieval algorithm based on local auto color correlograms is designed. According to the feature templates of flame and heavy smoke, the images included in explosion scenes are retrieved. The experimental results show that both coverage rate and accuracy rate of the image retrieval algorithm based on auto color correlograms are more than about 70%.Finally, the application of audio technologies in the violence scenes retrieval is researched. Audio data, audio frame features and audio section features are analysized. Since it is difficult to retrieve shooting scene by the features of image, the algorithm based on ZCR-STE is designed, in which a new method for calculating the short time energy (STE) of audio section is given. The experimental results show that ZCR-STE algorithm can retrieve the shooting scene effectively, and the new method for calculating STE of audio section is better than the average method.