Video Classification Based On Multi-features

With the rapid development of computer and multimedia technique, it is possible to make and store digital videos easily, so the digital videos become an appropriate source of information for various users like researchers. Additionally, Current information and communication technologies provide the infrastructure to transport bits anywhere. But it is difficult for users to search the videos in which they are interested from the mountains of video databases. In other words, many of these videos recording data are currently hardly usable, and this is mainly due to the lack of appropriate techniques, which can make the video content more accessible. So with both the rapid increase in the amount of generated video data and the wide range of video applications, an efficient and effective management of video records is much demanded. Manually indexing video content is currently the most accurate method, however, it is a very time-consuming process. For an user to retrieve the required information, automatic classification and categorization of the video content is essential.This dissertation makes a deep research on video automatic classification including analysis the state of the art, and summarizes the progress trend and drawbacks or advantages of these existing algorithms. Then we propose a new algorithm about video automatic classification based on various features and support vector machine (SVM) based on active relevance feedback.Considering that the average intensity, motion intensity and color distribution are different among regions of video, we propose a novel feature extracted method based on video regions. First, it divides the video into blocks, then according to comparison of the average intensity among different blocks of key frames to get the feature block intensity comparison code (BICC), and get the block color histogram through the statistics of color components in each block and extract the texture of key frames. Then extract the motion intensity of every region of videos. Furthermore, using principal component analysis (PCA), the extracted features are reduced the redundancy while exploiting the correlations between the feature elements. Finally, we design a tree classification model based on SVM with active relevance feedback, and use it to classify the videos with extracted features. The experimental results show that the proposed approach in the dissertation outperforms other methods which are based on features such as video saliency regions or only BICC.