Research On Key Technology Of Content-based Video Retrieval

Facing diverse and complex video data,it is a very import problem thathow to browse and retrieve video information quickly.Due to the traditional text-based information retrieval technology has been difficult to adapt to the new demands in the video field nowadays,therefore,content-based video retrieval technology emerges.The key technologies involved in the retrieval process are: shot boundary detection,key frame extraction,feature extraction,feature matching and so on.This paper mainly studies the shot boundary detection and key frame extraction.In the aspect of shot boundary detection,the existing algorithms have the following defects:(1)when extracting the features of the shot,only a single feature is extracted,and main content of the image can not be well represented;(2)the video within flash easily confused with the abrupt shot and the camera /shot movement easily confused with the gradual shot;(3)different video needs to be based on experience to manually set the threshold.In the key frame extraction,there are some problems:(1)determination of the number of key frames;(2)redundancy of image frames;(3)convergence rate of key frame extraction algorithm.In view of the shortcomings of shot boundary detectionand key frame extraction algorithms,two improvements are proposed in this paper:First,a shot detection algorithm of self-adaptive dual thresholds based on multi-feature fusion is proposed.The algorithm first divides non-uniformly the video frames and extracts HSV color characteristic and LBP texture feature for each block.Then two kinds of features are fused to represent the main contents,and the connecting block feature vector into a composite feature vector as the feature of the whole image according to the weight coefficient of each block.Finally,abrupt shot and gradual shot in the video are detected by determining the two threshold adaptively.In this way,it can effectively distinguish the video flash and the abrupt shot and the camera / lens object motion and the gradual shot,which can improve the efficiency of shot boundary detection.The second aspect is a key frame extraction algorithm based on improved clustering method is proposed by combining with the need of human preset threshold and redundancy etc..Firstly,the hierarchical clustering method is used to obtain the initial clustering results,and then the artificial immune clustering method is used to optimize the clustering results,finally,the corresponding number of key frames are extracted.This method does not need to artificial set the number of cluster centers and can get the clustering results more stable in the process of clustering,and the extracted key frames are well represented,effectively reduces the redundancy.Finally,the above two algorithms are analyzed through doing experiments,which proves the effectiveness of the algorithm.