Visual Hashing Based Algorithms On Tensor Decomposition And Non-negative Matrix Factorization

Multimedia information security is a cross-research topic in the field of multimedia and information security,which has attracted extensive attention of researchers in recent years.As an important multimedia information security technology,multimedia hashing has been successfully applied to multimedia information retrieval,multimedia content authentication and multimedia tampering detection.Multimedia hash is extracted from multimedia,based on the content of a concise representation.The multimedia hash algorithm can map arbitrary multimedia data into a series of short hash sequences.In the practical application,the hash sequence is used to represent multimedia itself,which can effectively reduce the storage cost of multimedia data and the complexity of multimedia analog computation.In general,multimedia hash must satisfy two basic conditions: robustness and uniqueness.Robustness refers to the fact that even though the two multimedia data is different,if they are visually identical,they should have the same or very similar hash.In other words,multimedia hashing requires the ability to counter normal digital operations,such as data compression and geometric transformation.The uniqueness requires that different media should have different hashes,which means that the distance between different multimedia hashes should be large enough.In addition,multimedia hashing may need additional features in some practical applications.For example,in multimedia content authentication,the extraction of multimedia hashing requires a key to control and multimedia hashing needs to be sensitive to content changes.In this paper,multimedia hashing algorithm is studied.The main object is the visual media in multimedia,namely image and video.In particular,two new visual hashing algorithms are designed using tensor decomposition(Tensor Decomposition,TD)and non-negative matrix decomposition(Non-negative Matrix Factorization,NMF).The first algorithm is based on the image hash algorithm based on TD.By constructing the tensor from the image,the tensor decomposition is applied to the image hashing extraction.The second algorithm is based on the discrete cosine transform(DCT)and the video hashing algorithm of NMF,and the video hashing sequence is extracted by combining DCT and NMF to achieve better classification performance.The main research results are summarized as follows.1.Image hashing algorithm based on tensor decomposition is proposed.Tensor is a kind of generalized high order matrix form.Currently,tensor decomposition has been successfully applied in many fields,including data mining,image analysis,signal processing and computer vision.In this paper,the image hash calculation be viewed as a tensor extracted a compact,says a new image hash algorithm based on tensor decomposition,referred to as TD hashing.In order to improve the robustness of TD hashing,first from the normalized image construct a stable 3 order tensor,then use a named Tucker decomposition tensor decomposition approach to 3 order tensor decomposition into one core tensor and three orthogonal factor matrix.Since the factor matrix can reflect the intrinsic structure of the original tensor,the TD hashing uses the factor matrix to construct the hash sequence to ensure that the algorithm has a good uniqueness.Select 14551 image as the experimental data to test the algorithm performance,receiver operating characteristic(ROC)curve contrast experiment results show that the classification of TD hashing performance and the hash length of hash algorithm is superior to a variety of images.2.Video hashing algorithm based on DCT and NMF is proposed.In this paper,we use DCT and NMF to extract video hashing sequence and design a new video hashing algorithm.Pre-processing,the algorithm for video input first get normalized video,then video frames are grouped,grouping for each frame,frame data were extracted with random block strategy and DCT coefficient was calculated by DCT coefficient matrix to construct the frame of grouping characteristics,finally using NMF learning short signature from the characteristic matrix,grouping series all the signature of the final video can generate the hash.By using the video data in 2050 to verify the performance of hashing algorithm,the results show that the video hashing algorithm proposed in this paper has a good uniqueness to the common digital operation robustness.The comparison of ROC curves shows that the algorithm is superior to video hashing algorithm of multiple literatures in robustness and uniqueness.