Research On Video Retrieval Technology Based On Multiple Feature Fusion

With the rapid development of network technology and multimedia technology,multimedia has become one of the most important means for people to obtain information on a daily basis,and among them,video is welcomed by most people because of its integrated visual and auditory characteristics.With the rapid growth of Internet video data,how to quickly and accurately find the information you want in massive video data has become a big challenge.At present,the traditional method of manually labeling videos by humans has been unable to cope with the increasing number of videos,and Content-based video retrieval technology has emerged as the times require.The main steps of video retrieval based on multi-feature fusion are divided into:shot boundary detection,shot key frame extraction,image similarity matching and other technologies.This paper mainly studies three aspects of shot boundary detection and shot key frame extraction and video retrieval.At present,in the aspect of shot boundary detection,the existing algorithms mainly have the following defects:(1)Extracting a single feature cannot fully express the video content.(2)A balance cannot be struck between efficiency and accuracy.In terms of key frame extraction,the existing algorithms are mainly clustering algorithms.However,the clustering algorithm needs to manually set the initial clustering center and the number of clusters.This may not be specific for different types of videos.In terms of video retrieval,the accuracy of existing algorithms in feature extraction and video similarity measurement still needs to be improved.For shot boundary extraction,shot key frame extraction.video retrieval technology,the following work is done in this paper:(1)This paper proposes a shot segmentation algorithm based on SURF and SIFT features.The algorithm is divided into two steps:initial inspection and re-inspection.In the initial inspection,first extract the HSV color histogram of the frame image.Judging the cut shot of the adjacent frame.Then extract the SURF features and combine the color features as the overall features of the image,perform cut frame re-judgment on the adjacent frames,according to the different characteristics of the transition time of the cut frame and the gradual frame,by using double thresholds combined with improved method of sliding window combination is used to obtain candidate frame of shot boundary.Finally,the obtained candidate frame set is rechecked by using SIFT feature combined with double threshold and sliding window technology,and finally the final shot transformation frame is obtained.This method of lens boundary detection not only makes use of the high efficiency of SURF features,but also takes advantage of SIFT features to improve the accuracy of lens boundary detection.Experimental data shows that this algorithm can effectively identify and distinguish between cut and gradual shots,and its accuracy is higher than other algorithms.(2)A key frame extraction algorithm based on improved K-Means clustering algorithm is proposed,based on the idea of cell phagocytosis.Using image entropy as the basis for clustering,Bottom-up of each key frame,and continuously merge the class of each key frame,the condition of the merge is based on the idea of k-neighbor.Finally,in the resulting class,use the the boundary of shots method collects frames at a fixed frequency as key frames.This method does not need to manually set the initial clustering center and the number of clusters,and the extracted key frames are highly representative.Experimental data shows that this method solves the shortcomings of K-Means,which requires manually setting the initial clustering center and the number of clusters,and the extracted key frames are highly representative,and the accuracy of this method is higher than the other algorithm.(3)According to the shot segmentation algorithm and the key frame extraction algorithm mentioned above,a video retrieval method based on the fusion of color features and SURF features is implemented.The image to be searched and the image of key frames that have been extracted are divided into blocks.The color histogram distance of each block is compared and scored using a Gaussian window function.The sum of the scores is assigned a weight of 40%,and then the SURF feature is used to extract the matching points of the two images,calculate the bad match points,assign a 60%weight,and add the two scores to get the total score.Query all videos in the database,divide the key frames of each video into overlapping subsequences,calculate the distance from the measured image and take the average as the score.The video with the lowest score is the video to be searched.Experiments show that this method,through the fusion of multiple features and the use of sliding windows,makes the average accuracy of the algorithm in this paper is 96%,which is higher than the algorithm in the literature and has a higher accuracy.In summary,the experimental results prove that the algorithm in this paper can accurately and efficiently extract the shot boundaries and key frames of video files and achieve accurate video retrieval based on this.