Research On Key Technologies Of Content-based Video Retrieval

In recent years,thanks to the popularization of 4G technology,the maturity of 5G communication technology and the improvement of mobile smart devices,video data has grown rapidly.How to quickly retrieve videos that users are interested in from massive video databases has become a meaningful topic in information age.Traditional video retrieval methods based on text have been unable to meet the increasing needs of users.Therefore,content-based video retrieval methods have emerged.This thesis conducted an in-depth study of the key technologies of content-based video retrieval,including key-frame extraction,feature extraction and representation.In terms of key-frame extraction,the existing algorithms have the following shortcomings: The number of key-frames is limited to a fixed value,and the extracted key-frames cannot fully describe the content of complex video shots.Classical method is not sensitive to the motion of the camera,which leads to unstable performance of key-frame extraction.In terms of feature extraction and representation,there are the following shortcomings: traditional encoding methods lose a lot of information when representing high-dimensional features,resulting in a decrease in retrieval accuracy.Convolutional Neural Network Hashing cannot extract features and represent hash features at the same time.Aiming at the above disadvantages,an efficient content-based video retrieval algorithm has been proposed.The main contributions are as follows:In terms of key frame extraction,we made improvements on the traditional cluster-based key-frame extraction algorithm,which can automatically determine the number of key-frames based on the video content.Experiments were carried out on the datasets commonly used in video tasks and video clips published on the Internet.The experimental results show that,in terms of compression rate and fidelity,the adaptive key-frame extraction algorithm we proposed performs significantly better than other key-frame extraction algorithms.As far as video feature extraction and representation,we designed a feature representation method based on pooling to reduce the dimensionality of high-dimensional convolutional features.Ordinary pooling is to sample the feature map on the same feature channel,which results in more information loss.The feature representation we proposed is to sample the feature maps of different channels to minimize the loss of information.The experimental results show that the pooling feature method has certain advantages in retrieval accuracy.In order to further improve the retrieval speed,a feature extraction and representation method based on deep hash model has been proposed.This model combines convolutional neural network and hash technology,which can directly extract the hash features of video frames after training.The experimental results show that when the feature dimensions are the same,the deep hash model we proposed is significantly better than other hash methods in terms of retrieval accuracy,and the retrieval speed is also significantly improved.