Support Vector Machine based multi-View hashing approach for Near Duplicate Video Retrieval

Abstract

Near duplicate videos (NDVs) are the primary concern for cloud storage and web search. Similar video-sharing triggers issues related to copyright and financial loss to the maker of the video. We propose a near-duplicate video retrieval (NDVR) system. The proposed algorithm is trained and tested on the benchmark standard dataset CC-WEB-VIDEO. For video processing, we first split it into frames. We have processed 48 GB of frames retrieved from 80GB of video dataset. Hue Saturation Value (HSV) and Local Binary Pattern (LBP) are used to capture the global and local features of frames. It is observed that 3% to 10% of frames have similar frames. Therefore, the kernel-based component analysis (KPCA) algorithm is used to reduce the redundant frames. A deep Convolution Network-based VGG16 algorithm is also used to identify the best strategy for NDV. Finally, the feature extracted from all three techniques, HSV-KPCA, LBP-KPCA, and VGG16, are trained and tested on a radial basis function-based support vector machine (RBF-SVM) classifier. RBF is used to address the non-linearity of nonredundant frames. Results are compared with state-of-the-art algorithms for NDVR. The proposed system reports a higher Mean Average Precision (MAP), Area under the curve (AUC), and accuracy than previous NDVR systems.