Study On Algorithm Of Content-Based Video Retrieval

With the rapid development of Internet and computer technology, a large number of image and video data is being produced, and stored on the Internet or the local computer, in people's daily life and work. Facing with the massive multimedia image and video resources, users wish to retrieval their interested target quickly and precisely.The method of content based video retrieval find out the interested image, according to the characteristics of query image, in massive video sequence frames. Digital image processing and pattern recognition technology have been utilized for video key frame extraction, and image retrieval based on scale invariant feature in salient regions. The following three aspects are mainly researched in this paper:(1) A key frame extraction method based on ReliefF feature-weighted and fuzzy c-means algorithm is researched. Clustering algorithm is used to cluster similar frames with visual characteristics, and representative samples could be selected as key frames. Aiming at fast changing video shot content resources, initial key frames are obtained based on fuzzy c-means algorithm firstly, ReliefF algorithm is employed to assess the feature weight vector of each category, fuzzy c-means clustering is performed using the weighted image features secondly, frames have the largest membership are selected to represent the whole categories.(2) A key frame extraction method based on hierarchical AP clustering is researched. Since traditional AP algorithm is inappropriate to the large-scale pictures clustering, the video frames set is divided into several subsets firstly, and the traditional AP is used to obtain the exemplars of each subset. Then the adaptive AP is implemented on the obtained exemplars secondly. The key frames of video sequences are extracted finally, according to the index of Silhouette for the best clustering results. Exemplars of categories are selected as key frames.(3) An image retrieval method based on key points in saliency regions is researched. Invariant local points are extracted for matching specific targets and corresponding images in video retrieval. Since there are too many key points in one image to match similar images, and complex background would influence result of matching, vision saliency is introduced. Saliency map is utilized to detect subject regions, extraction and matching of key points are competed in salient regions for image retrieval. This method can eliminate the interference of the background effectively.The experimental results show that proposed methods for key frames extraction and image retrieval are useful and efficient. Extracted key frames reflect the main content of the video accurately, the execution speed of algorithm is fast, and also have a high compression ratio, image retrieval algorithm can reduce the impact of background factors, and has strong robustness at the same time.