Binocular Depth Feature-based Visual Information Retrieval Research

Traditional retrieval of information is mostly based on manual annotation, whichnot only wastes a lot of human resources, but also can not meet the requirement of theobjectivity of retrieval. So, technology of content-based information retrieval is putforward accordingly, its research is of great significance.With the analysis on traditional content-based video information retrieval, thisstudy uses the theory of binocular vision, mainly consisting of the pole geometry, stereomatching and other relevant technology and knowledge, proposes that the depth featurecan be used in video retrieval, which make up the shortage of the information such ascolor, shape, texture etc. This method transforms two-dimensional video information tothree dimensional space, making the video’s features richer, and bringing betterperformance in retrieval. At the same time, combining with the video sequencecharacteristics, sequence pattern matching technology based on depth is proposed，theapplication of the depth sequential patterns rather than frame’s depth information invideo matching, makes lower matching cost and better efficiency, which shows wellrobustness in the process of retrieval，besides，using the depth sequential patterns notonly increases the reliability of video matching, but also overcome the limits of differentlengths of different videos in retrieval, which makes the application of video retrievalmore flexible. In addition, in the aspect of building video library, this thesis proposedthe idea to use index tree based on prefix tree, and index tree improved on this basis,index tree reduces the space requirements of video database effectively, and inherits theprefix tree’s features, by which one particular feature can position at all related video.The advanced index tree can give an evaluation to video match degree in the process ofretrieval, making the returned results in correct order by the matching degree whileguaranteeing the Accuracy.Finally, this study uses the public network video library named CC_WEB_VIDEOin the experiment, compares the main experimental procedure including shot detection,depth information extraction, retrieval with other relevant methods and gives analysis,achieves higher accuracy and recall rate in the result.