Study Of Video Retrieval Method Based On Hashing Learning

In recent years,with the rapid development of the Internet,there are lots of video data on the Internet,so that video retrieval algorithms have also encountered many new problems.In order to efficiently and accurately retrieve target videos,video hashing algorithms have received more and more attention.However,most existing supervised video hashing algorithms design hash functions based on pairwise similarity or triple-wise relations and focus on local information,which leads to low learning efficiency of the algorithms.From a global perspective,the algorithm in this paper encourages similar videos to converge to the same hash code,and dissimilar videos are mapped to different binary codes to generate more discriminative hash codes.The main work of this paper is introduced as follows:(1)A supervised video retrieval algorithm based on discriminative codebook hashing is proposed,which solves the problem that current video hashing algorithms ignore the global structure.The algorithm learns a more discriminative codebook according to the label information of the video.Similar keyframes share the same codeword,while dissimilar keyframes have different codewords.Specifically,the construction of the codebook firstly controls the element distribution of the same bits of different codewords through Bernoulli Distribution.Then,the distance between the two codewords is guaranteed to be as maximal as possible through constraints.Then the composite Kullback-Leibler Divergence is used to save the similarity structure of the original space into Hamming Space.Finally,the gradient descent algorithm is used to optimize the solution.The experimental results show that when the hash code length is 32 bits,the mean Average Precision on the datasets CC＿WEB＿VIDEO,HMDB51,and UCF101 are 0.9531,0.2568,and 0.5310,respectively.(2)A fast video hashing algorithm based on similarity preservation and discriminant analysis is proposed,which solves the problems of long training time and low accuracy of current video hashing algorithms.The algorithm directly obtains the similarity matrix through the label information of the video,which greatly reduces the calculation time of the similarity between the video key frames.At the same time,the codebook matrix defined according to the label information is used,so that the hash codes generated by different categories of videos are more different.In the process of algorithm optimization,the method of iterative optimization is used to find the closed solution of the parameters through the precomputed intermediate term,and learn to obtain the video hash function.The algorithm makes the similarity matrix and the codebook matrix work together,and introduces a precalculated intermediate term,which not only obtains a high accuracy rate,but also greatly shortens the training time of the algorithm.The experimental results show that when the hash code length is 32 bits,the mean Average Precision on the datasets HMDB51 and UCF101 are 0.4079 and 0.7710,respectively.