Research On Near-duplicate Video Detection Based On Correlation Analysis

Near-duplicate video detection has come into sight with its profound prospects in video surveillance and tracking,copyright violation detection,video database cleaning and video recommendation.How to accurately and quickly get the required video from massive video has become a hot research direction of video.Based on improving the overall performance of near-repeated video detection,this thesis proposes a model algorithm applied to near-repeated video detection for requirements of accuracy and rate.The main tasks are as follows:1.In terms of video detection accuracy,a Kernelized Canonical Correlation Analysis Ensemble with Laplacian Regularization(LRKCCE)method was proposed.Multiple kernel canonical correlation analysis models were used in the integration framework to reduce the total least squares loss in the multiple kernel Hilbert space(RKHS).The multi-core model enables a single KCCA with better correlation between data to be automatically assigned a greater weight,reducing the overall sensitivity of the algorithm to different kernel functions and parameter selection.At the same time,the new RKHS composed of multiple different high-dimensional feature spaces can more effectively use different mapping methods of different kernel functions,so as to achieve a better combined function mapping effect.The experiments carried out on the remote sensing dataset and near-repeated video dataset NWPU-RESISC45 show the performance of the method.2.In terms of video detection efficiency,Deep Canonical Correlation Analysis based on Fast Fourier Transform(FFT-DCCA)is proposed.The Research on image detection of fast near-duplicate content based on depth method and cyclic matrix.Deep Canonical Correlation Analysis(DCCA)can effectively improve the robust results of CCA method in near-duplicate video detection.At the same time,because of the need of quickly detect near-duplicate images in massive images,fast Fourier transform based on cyclic matrix will also be developed.The proposed method can make the operation element-by-element processing by using the FFT,so as to obtain afast near-duplicate video detection method.The results of the experiments on the CC_WEB_VIDEO dataset show that the algorithm can obtain better near-duplicate video detection results in terms of accuracy and speed than any other algorithm.3.Complete a near-repetitive video detection system with FFT-DCCA algorithm as the basic algorithm model.The system is divided into two modules,the near-duplicate video detection function is friendly to all users.