Research On Key Technologies Of Video Retrieval Based On Convolutional Neural Network

In recent years,with the development of Internet technology,a large amount of video information has appeared in people's lives,and video retrieval has gradually become an important way for users to obtain information.The video contains rich information,and with the popularization of high-definition,the video data is increasingly huge.How to quickly and accurately retrieve the massive video data generated on the Internet is an important issue facing the development of the video industry.The traditional video retrieval based on text annotation is to manually label the video with text,which is inefficient and subjective,and generally cannot accurately describe the meaning of the video;the traditional content-based video retrieval uses a convolutional neural network Extract the lower-level feature information of the image to establish an index,and implement similarity retrieval of video feature vectors according to a certain similarity measurement algorithm.Although the accuracy is high,when faced with a large-scale video,the calculation is large and the efficiency is low.This article combines the above two methods,according to the characteristics of sports video,through the improved GoogLeNet network to extract the underlying video features to train the model and tag the video,and finally the proposed keyword-based fuzzy search algorithm based on ontology semantic expansion to tag words Perform a search.It not only ensures the accuracy and objectivity of the label,but also improves the retrieval efficiency.In this paper,by studying the characteristics of sports video,a histogram difference method based on automatic threshold and a four-step method based on block matching are proposed to perform mutation detection and gradient detection of video shots,respectively.Through the adaptive threshold,the area with large frame difference changes is marked as the candidate area of the shot,and then the shot boundary is determined by the mutation detection algorithm and the gradient detection algorithm respectively.A key frame extraction algorithm based on clustering and optical flow analysis is proposed to extract key frames of sports video.The algorithm effectively removes redundant frames,and the extracted key frames are also more representative.Through a lot of experiments,the keyword-based fuzzy search algorithm based on improved GoogLeNet network and ontology semantic extension proposed in thispaper shows ideal retrieval performance.It is feasible to use this method to extract,annotate and search for the underlying features of videos And a prominent feature of the algorithm is that it can quickly and efficiently retrieve the required video from a large number of Internet video resources,while reducing the false detection rate and the missed detection rate,while improving fidelity,basically meeting people's daily needs.