Research On The Technology Of Content-based Re-ranking Video Retrieval

With the development of the Internet and the computer technology, the video data isincreasing explosively in the Internet. To meet the demands that the network users caneasily obtain the expected information, Web Search Engines have been developed contin-uously. However, popular Web Search Engines are mainly based on the method of textindex to obtain the video information annotated by text. Apparently, the simple text anno-tation can’t provide the accurate and specific description for a video which includes plentyof multimedia information. Then, this makes it is urgent to research the content-basedvideo retrieval. Many researchers have focused on how to utilize the image and audio in-formation to increase the retrieval precision.In this paper, the task of video retrieval is divided into two parts according to theprocess of retrieval analysis, including video analysis and retrieval framework. Videoanalysis mainly includes the processing of video structure and the information extraction.In video analysis, shot segmentation and key-frame description are main research prob-lems in this paper. In the retrieval framework, classification algorithm and re-rankingtechnology will be researched to refine the relevance of query results.A video clip can be divided into scene, shot and frame according to the video structure.Shot is often a minimal unit for video processing. Thus, in the first step of video retrieval,a video clip will be segmented several shots. In this paper, a fast coarse-to-fine video shotsegmentation algorithm is proposed through analyzing the deficiency of current algo-rithms about the precision and the efficiency. The camera/object movement and gradualshot change can be differentiated through this method. Based on the improved informationentropy theory proposed in this paper, the differences between the shots of a video se-quence are calculated. The adaptive thresholds are implemented to select the sequences of candidate shots from the video sequence. Because the camera/object movement and thegradual shot change present similar characteristics, they are all selected as the candidateshots. Then a fast moving-edge-detection algorithm proposed in this paper is implementedto distinguish the gradual shot change. The fast coarse-to-fine video shot segmentationalgorithm proposed in this paper is based on the statistical properties of the characteristics,thus the computational complexity is reduced effectively.In the stage of the shot description, the dynamic and static features are used to describea video shot. The dynamic descriptors are established by analyzing the motion changebetween the consecutive frames. The dynamic description is used as an auxiliary methodof video retrieval in this paper. The key-frame technology is used as the method of thestatic description. In the stage of the key-frame description, the defect of local feature isanalyzed. The local feature descriptors are mainly based on the gray images, and colorcan’t be utilized. First of all, the defect of Shadow-shading Quasi-invariant is presentedand analyzed in color image, and the improved method is proposed. Then, the stable in-variant regions are obtained. The experiment of the repeat rate shows the improved meth-od increases the stability of the region detection. On the basis of the stable regions, thecolor invariant region descriptors are extraction. Meanwhile, the color invariant regiondescriptors and the local features are combined to form color Bag of Visual Words model(BOW). In the experiment, color BOW is compared with other similar BOW model. Theresults of the experiment show color BOW obviously increases the performance, becausethe color clue is added into BOW.In the framework establishment of video retrieval, the re-ranking algorithm based onmain-class is proposed, and the structured processing of the video, the visual words, andthe re-ranking algorithm are combined to establish the retrieval framework. In the processof video retrieval, the initial results returned by text-based search engine are clustered, sothat the class of the results is annotated automatically. Then, the linear discriminant modelbased on main-class is proposed by analyzing the feature space of the samples. Finally, the proposed video structured method, shot descriptors based on visual words, and re-rankingmethod are combined to form our retrieval framework, and the precision of video retrievalare increased in the final experiment.