An approach to content-based video retrieval

We propose an approach to retrieve videos similar to the given query video from the database. Our proposed approach consists of three phases. First, every database video is segmented into several shots. Second, for each shot, one or more key frames are selected, and then a feature vector for each key frame is computed. So, every database video is transformed into a sequence of feature vectors. Third, to retrieve videos similar to the given query video, the query video is also transformed into a sequence of feature vectors. Then, we use a dynamic programming approach to compute the similarity between the query video and each database video. The database videos with similarity higher than a predefined threshold are output and returned to the user The experimental results demonstrate that the more key frames that are used in the query video, the better the precision and recall that can be obtained.