Video Copy Detection Technology Based On Tensor Decompositions

In the age of the Internet, the development of terminal equipment and mobile internet enable users to access and publish videos conveniently. Meanwhile, video processing software facilitates the edition of videos including changing the format of encoding, the brightness and contrast ratio, scaling, adding logo, etc. These convenient tools not only enrich people’s cultural and entertainment life, but also breed more and more identical or similar video copies. For better video retrieval efficiency and consumption environment, video copy detection technology takes the wave and gradually becomes a hot topic of multimedia information study. Video hash, the core technology of video copy detection, extracts features of video content, and maps the content feature to a simple binary hash sequence. Video hash is the identifier of video content. It tells a copy from the original video by measuring the distance between different hash sequences.This paper introduces the basic knowledge of the video copy detection, video hash as well as the evaluation criteria. A video copy detection algorithm based on feature fusion and Manhattan quantization will also be presented in this paper. The main contributions are as follows:(1) The high-order tensor model is applied to acquire the comprehensive expression of a video on the basis of its global, local and temporal features. Differing from the currently existing single feature-based and multiple feature-based methods, the method proposed in this paper makes better use of the relationship between the features of video. For a better representation of the content, the global, local and temporal features are reflected respectively by the gray histogram, SURF(Speeded-Up Robust Features) and normalized pixel difference between adjacent frames. The video is represented as a third-order tensor, and the three first-order tensors are obtained as the fused feature of the video by tensor decompositions.(2) Manhattan Quantization method is applied in hash learning. Compared with the traditional Euclidean distance and Hamming distance, the distance measurement of Manhattan quantization can better reflect the structural similarity between the original spatial points and hash points. In addition, the traditional hash quantization methods encode all the dimensions of original features without considering the differences among different dimensions of them, and destroy the structural similarity between the original features. The proposed method encodes each dimension of original features respectively and maintains good structural similarity.