Research And Implementation On Content-based Video Shot Retrieval Method

In the Internet era, thousands of images, video and other multimedia information generated every day. In the face of large-scale multimedia data, the existing methods are often based on keyword search, such as Google, Bing or Baidu, etc. These methods mainly utilize metadata, the information of the text and other information around the multimedia. They still base on text and the retrieval results may not related to the content. For many multimedia data don't have text labels, these retrieval methods have been unable to meet users' requirements of big data retrieval in the era. Content based video retrieval has gradually become one of the most important issues of research and application.Generally, content-based video retrieval includes video key frame extraction, video feature extraction and similarity measure. The key frame extraction is relatively mature, so in this paper we mainly research and analys the video feature extraction and similarity measure.In the feature extraction, we put forward a retrieval method based on the bow features of video shot, do a comparative analysis of the key influence factors point detection and feature dimensions of bow features using the Harris Laplace and Hessian affine detectors and SIFT descriptor, construct bow features of different dimensions with the bag of words model, and normalize to obtain the final comprehensive video feature representation. This method works better than those with single feature.In similarity measurement, we use the SVM Multibag similarity measure method, through a combination of multiple SVM classification model, make full use of the user samples in the retrieval process and the negative samples in the data. After the retrieval result the paper further establishes a semi supervised learning model to reorder for the final results. The experimental results show that Multibag SVM can achieve better effect than the traditional direct distance metric, and semi supervised learning can improve retrieval effectiveness.Finally we output a set of interactive video camera retrieval application with GUI, to present directly to the user the video camera results retrieved in our method.