Local Features Based Image Sequence Retrieval

We propose an approach to retrieve image sequences similar to the given query image sequence from database. Our proposed approach consists of three phases. First, the query image sequence and every database image sequence are segmented into several shots based on histogram difference between local images of consecutive frames. Second, for each shot, one or more key frames are selected based on histogram difference between local images of benchmark frame and each followed frame. Third, to retrieve image sequences similar to the given query image sequence, similarity between the key frames of query image sequence shot and key frames of each database image sequence is computed. The similarity is also measured using the histogram difference between local images of key frames. Local image is defined by interest points, that is, the regions around all interest points for a frame. The database video shots with similarity higher than a predefined threshold are output and returned to the user. Experimental results show that the proposed video shot detection method can overcome the deficiency of traditional histogram-based method that different contents frames have similar histograms, can distinguish between gradual changes and camera motions effectively, and can effectively detect both abrupt transitions and gradual transitions; the key frames selected by the proposed method have good representative power and can improve the performance of video shot retrieval.