Research And Implement Of Near-Duplicate Video Retrieval Based On Toeplitz PLS

Nowadays,with the technological development,the number and variety of Internet information is increasing,information explosion and fusion is leading the new era of Internet.Therefore,large-scale Near-Duplicate Video Retrieval(NDVR)is becoming more and more important.NDVR is a hotspot in video retrieval domain.It has come into sight with its profound prospects in video surveillance and tracking,copyright violation detection,video database cleaning and video recommendation.Aiming at the application requirements of NDVR for large-scale video databases,this thesis studies the method of NDVR with high accuracy and fastness.The main work is as follows:1.A Toeplitz based Kernel Partial Least Squares(TKPLS)algorithm is proposed.TKPLS introduces kernelization technique into Partial Least Squares(PLS)to seek a latent nonlinear correlation between videos in NDVR.Kernelization can extend linear to nonlinear analysis by projecting data to high dimensional space but on the contrary causes computation cost problem.To solve the high-dimensional problem brought by kernelization,the Toeplitz circulant matrix is applied in Kernel Partial Least Squares(KPLS)to form the TKPLS algorithm.This optimizes the kernelization process to element-wise operation based on the property of Fast Fourier Transform(FFT).TKPLS can achieve low computational complexity and nonlinear representation in NDVR task.The experimental results on the CC_WEB_VIDEO dataset show that TKPLS can obtain better Near-Duplicate Video Retrieval effects in terms of accuracy and speed than BCS,SE,SSBelt,and CCA algorithms.2.A Toeplitz based Multiple Kernel Partial Least Squares(MTKPLS)algorithm is proposed.The idea of Multiple Kernel Learning(MKL)is applied to the proposed TKPLS algorithm.A multiple kernel weight is used to linearly combine multiple kernel functions to transform a single kernel model into a multiple kernel one.The sensitivity of parameters in kernel techniques is naturally solved since the proposed can automatically adjust through multiple kernel weights learning.Further,multiple kernel mapping combines high-dimensional projection spaces into ensemble Hilbert feature space which makes full use of various kernel functions capabilities.The experimental results on the CC_WEB_VIDEO dataset show that the MTKPLS algorithm is more efficient and accurate than the BCS,SE,SSBelt,CCA,and TKPLS algorithms in terms of accuracy and speed.3.Based on the MTKPLS algorithm,a NDVR system is designed and implemented.The system is divided into user management module and NDVR module.A performance test run of the system shows that,the system interface is user friendly and well-functioning.The accuracy and fast retrieval results verify the effectiveness of the proposed method.