Video Hashing Algorithms Based On Multidimensional Scaling And Invariant Moments

As an important research topic in the field of digital media content security,video hashing has attracted extensive attention from researchers in recent years.It has been successfully applied to many applications,such as video copy detection,video authentication,video retrieval and video tampering detection,etc.Video hashing algorithms can extract a content-based and short sequence of digits or bits from an input video.The calculated sequence is called the video hash of the input video.In practical applications,the video hash can be used to represent the video itself,which can effectively reduce the storage cost of video and the complexity of video similarity calculation.Generally,video hashing algorithms should satisfy two basic properties,namely robustness and discrimination.Robustness means that visually identical or similar videos should be mapped to the same or similar video hashes.After undergoing digital operations such as brightness adjustment and contrast adjustment,the specific data representation of the video has changed,but the visual content is still similar to the original video.Therefore,this property can ensure that the video hashing algorithms can correctly identify similar videos after undergoing normal digital operations.Discrimination means that videos with different visual contents should have different hash values.This property can ensure that the video hashing algorithms can correctly distinguish videos with different contents.Because the above two properties are restricted,it is an important task to design a video hashing algorithm with good classification performance between robustness and discrimination in current research of video hashing.In this paper,video hashing algorithms are investigated by the theories and technologies of multidimensional scaling(MDS),ordinal measures,three-dimensional discrete wavelet transform(3D-DWT)and invariant moments.And two novel video hashing algorithms are proposed.The first video hashing algorithm is based on MDS and ordinal measures,the second video hashing algorithm is based on 3D-DWT and invariant moments.The main research results are summarized as follows:1.A video hashing algorithm based on multidimensional scaling and ordinal measures is proposedMultidimensional scaling is an effective technique of data dimensionality reduction,which can learn low-dimensional features from high-dimensional data.It has been successfully applied to many applications,such as object retrieval,object localization,data visualization and image hashing.Ordinal measures are features which have good robustness and discrimination and have been applied in the fields of video and image processing.In this paper,we propose a video hashing algorithm which combines multidimensional scaling and ordinal measures.Specifically,the input video is preprocessed by temporal-spatial resampling and color space conversion.Then the input video is divided into several video groups and secondary frames are extracted from video groups.Next,one level 2D-DWT and the mean of gradient magnitudes are used to construct highdimensional feature matrix.After that,multidimensional scaling is exploited to learn lowdimensional features from high-dimensional feature matrix.Finally,the technique of ordinal measures is used to quantify the features.The performance of the algorithm is tested on an open database.Experimental results show that this algorithm is robust against many digital operations and has good discrimination.2.A video hashing algorithm based on 3D-DWT and invariant moments is proposedThe 3D-DWT is a useful technology of video processing and robust to some digital operations such as noise and compression.It can achieve initial data compression.The invariant moments have good robustness and discrimination,and have been widely used in the field of image processing.In this paper,the 3D-DWT and invariant moments are used to design a video hashing algorithm.The input video is preprocessed by temporal-spatial resampling and color space conversion.Then the input video is divided into several video groups.Next,the two-level 3DDWT is used for each video group and the coefficients of DWT are exploited to construct secondary frames of all video groups.And then the invariant moments are extracted from all secondary frames.Finally,invariant moments of all secondary frames are quantified to form video hash.Experimental results show that this video hashing algorithm is robust to many normal digital operations and has good discrimination.Receiver operating characteristic(ROC)curve is used to compare classification performance with several existing video hashing algorithms.Experimental results show that the proposed video hashing algorithms outperform the compared algorithms in classification performance.