Robust Image Hashing Based On LLE And NMF

Image hashing algorithm is an important research topic of image content information security.It has been widely used in image retrieval,image indexing,digital watermarking,image authentication and copy detection.Essentially,image hashing algorithm is a one-way mapping function,which maps any-size image into a fixed-size and short sequence of bits or digits.This bit/digit sequence is called image hash.In practice,image hash is used to represent the image itself and thus can effectively reduce the cost of image storage and the complexity of similarity computation.Generally,image hashing algorithm should meet two properties.(1)Robustness:This property means that if two images have the same visual content,hashing algorithm should map them to the same image hash or very similar image hashes no matter whether their digital representations are the same or not.This property ensures that image hashing algorithm can correctly identify those similar image versions undergone content-preserving operations such as JPEG compression,watermarking embedding,rotation,brightness and contrast adjustment,scaling,gamma correction and Gaussian low-pass filtering.(2)Uniqueness:This property implies that,if two images have different visual content,image hashing algorithm should map them to different image hashes.This property ensures that image hashing algorithm can effectively distinguish different images.Except robustness and uniqueness,image hashing algorithm has special requirement in some practical applications.For example,the application of image authentication requires security.It means that image hash generation should be controlled by secret key and different keys should produce different hashes.This paper exploits the techniques of color vector angle(CVA),discrete cosine transform(DCT),locally linear embedding(LLE),singular value decomposition(SVD)and non-negative matrix factorization(NMF)to develop novel image hashing algorithms,and achieves two meaningful results,i.e.,robust image hashing algorithm via DCT and LLE and robust image hashing algorithm via SVD and NMF.The detailed research results are as follows.1.A robust image hashing via DCT and LLE is proposedCVA is sensitive to the change of hue and saturation,but it is not sensitive to the brightness adjustment.Compared with the Euclidean distance of the color,CVA is more effective in measuring color changes and therefore suitable to describe the feature of color image.Considering these advantages of CVA,I propose to exploit DCT and LLE to extract image hash from CVAs of the whole image.Specifically,bi-cubic interpolation is first used to convert input image to a normalized image with fixed size,and Gaussian low pass filter is used to process the normalized image.CVAs of the whole image is then extracted and further divided into non-overlapping blocks.Next,DCT is applied to each block and those low-frequent DCT coefficients of each block are taken to form feature matrix,whose columns are scrambled to achieve security.Finally,LLE is applied to the scrabled feature matrix,and those variances of LLE low-dimensional vectors are encrypted and quantized to make image hash.Experiments show that the proposed algorithm is robust against various normal digital processing,such as JPEG compression,watermark embedding,brightness and contrast adjustment,scaling,gamma correction and Gauss low-pass filtering,and achieves good uniqueness.2.A robust image hashing via SVD and NMF is proposedSingular value of image is generally stable.When an image is slightly disturbed by digial operation,its singular values will not be significantly changed.Considering this,I proposed to design image hashing algorithm by combining SVD and NMF.Specifically,bi-cubic interpolation is exploited to convert input image into a fixed-size image,and the fixed-size image is divided into non-overlapping blocks.SVD is then applied to each block and those singular values in the diagonal matrix of SVD results are used to form a secondary image.To enhance security,columns of the secondry image are randomly scrambled and the scrambled result is filtered by a Gauss low-pass filter.Finally,NMF is applied to the smooth secondary image and the ordinary measure of NMF coefficients is used to generate image hash.Experiments show that the proposed algorithm is robust against common digital processing,such as JPEG compression,watermark embedding,brightness and contrast adjustment,scaling,gamma correction and Gauss low-pass filtering,and has good uniqueness.Receiver operating characteristic(ROC)curve is used to analyze classification performance of the proposed hashing algorithms in robustness and uniqueness.ROC curve comparisions with several existing hashing algorithms are also discussed.Experimental results show that the proposed image hashing algorithms outperform the compared algorithms in classification performance.