Researches On Performance Optimization Of Hashing For Video Retrieval

With the rapid development of multimedia and Internet technology,video shows explosive growth as a media form with the most abundant content.Under these circumstances of large-scale video data,it is significant to achieve the fast retrieval of video content under the limitation of bandwidth and computation cost,which has not only profound meaning for information field,but also social and business value.Video hash is a process of turning high dimension video information into the binary,reducing the amount of calculation,memory requirements for video matching calculation and the bandwidth of data interaction,so as to realize the high speed of video information retrieval.Efficient video hash should have:1)accurately represent the features of video;2)a hash map that preserves robustness and differentiability;3)optimal hash length that can maintain the performance of hash.Therefore,the video hashing can be applied in large-scale video retrieval,only after the problems of video features,hash mapping and binary code length have been solved.This paper focuses on the performance optimization of hashing technology for video retrieval.Firstly,it starts with the research background and the state of art.Then the system composition and evaluation criteria of hashing for video retrieval are introduced.Then,based on the theory of rate distortion,the optimal hash code length is obtained on the basis of ensuring the robustness and differentiability of hashing.Finally,deep learning is applied to extract video features,and a new video hashing algorithm based on 3D-CNN is proposed.The main innovation and contribution of this paper are as following two aspects:(1)A hash length optimization algorithm is proposed.Although hash representation has attracted increasing attentions in recent years,hash length is still a neglected element in the evaluation of hashing.Hash length is the dimension of hash representation,which is important for the performance of video hash.We try to define the optimal hash length according to the probability of collision(PoC)of hash.Based on this definition,we demonstrate that this optimal hash length can be predicted from a small portion of dataset.By mathematical modeling of the bit error rate(BER)and probability of collision(PoC)of the test data,the optimal hash length is obtained on the basis of ensuring the robustness and differentiability of hashing.(2)A 3D-CNN based video hashing method is proposed.The video features adopted by traditional hashing algorithm are mainly "hand-crafted",which is based on the relevant knowledge design of video processing that the researchers already have.The extraction of such kind of features requires significant prior knowledge and the feature is usually low-level.So that a learning-based video feature is used,which is obtained via a 3D-CNN model.The 3DCNN-based features can represent both spatial and temporal information of videos,as 3D convolutions used in 3DCNN can capture the motion information through multiple adjacent frames.A video hashing algorithm is proposed based on the CNN-based feature,which is defined as CNNF.The CNNF has captured some high-level semantic characteristics of video,which is suitable for the large-scale video retrieval on video hashing.Based on video features,the deep network is utilized to obtain and optimize the features that can maintain the similarity of video.Starting from the perspective of the binary process,the hashing length optimization is studied to get optimal length on the basis of guaranteed differentiation.Through these targeted research strategy,the efficient video hash is obtained to meet the requirements for large-scale video search,to offer support for the theory study and practical application in the fields of public safety,video sites,mobile search.