Image Hashing Based On Multidimensional Scaling And Wavelet Statistical Features

Image hashing is a frontier research topic in the field of content security of digital media.It can map any size image into a short string of characters or digits,and has been widely used in image retrieval,watermark embedding,image tampering detection,image quality assessment,and so on.In practical application,images are often suffered some normal digital processing,such as JPEG compression,brightness adjustment and gamma correction.These operations will change digital representation of the image,but do not change visual content of the image.Therefore,image hashing should map these images with visually identical content into the same hash or very similar hashes.This is the first basic property of image hashing,called perceptual robustness.Another basic property of image hashing is discrimination,which requires that those images with different visual contents should have different hash sequences.Obviously,perceptual robustness and discrimination contradict with each other.In general,improvement of perceptual robustness will lead to decrease of discrimination and vice versa.Keeping a balance between robustness and uniqueness is an important index of image hashing research.This paper exploits the theories and techniques of multidimensional scaling,log-polar transformation,edge detection and wavelet transform to investigate image hashing algorithm,and obtains two meaningful results.The first research result is the robust image hashing with multidimensional scaling,which can resist any-angle rotation and has a good discrimination.The second result is the robust image hashing based on edge detection and wavelet statistical features,which can achieve a good balance between robustness and discrimination,and also can be applied to the reduced reference image quality assessment.The detailed research results are as follows.1.A robust image hashing with multidimensional scaling is proposedMultidimensional scaling(MDS)is an efficient technique for data analysis,and has been successfully applied in data visualization,object retrieval,data clustering,and so on.However,its use in image hashing is rarely discussed yet.In this study,we investigate the use of MDS in image hashing and propose an MDS-based hashing algorithm resistant to any-angle rotation.The proposed hashing algorithm extracts a rotation-invariant feature matrix with log-polar transform(LPT)and discrete Fourier transform from the normalized image,and learns a compact and discriminative representation from the feature matrix by MDS.The proposed hashing can resist any-angle rotation due to the following reasons.LPT can convert rotation in the Cartesian coordinates to translation and DFT has the property of translation invariant,which theoretically ensures anti-rotation capability of the proposed algorithm.Experiments show that the proposed algorithm is robust to many content-preserving operations,including speckle noise,salt and pepper noise,scaling,contrast adjustment,brightness adjustment,Gaussian low-pass filtering and any-angle rotation,and reaches good discrimination.Receiver operating characteristics(ROC)curve comparisons illustrate that our algorithm outperforms several existing algorithms in classification between robustness and discrimination.2.An image hashing with edge detection and wavelet statistical features is proposed.Human eye is the terminal of human vision system(HVS).When investigating image hashing for reduced reference image quality assessment,the HVS should be fully considered.Wavelet transform maps an image from spatial domain to frequency domain,and obtains various sub-bands representing different information of original image.The low-frequent sub-band is the coarse representation of the image and the high-frequent sub-band indicates detailed change of the image,such as edge,contour and texture.This is consistent with multi-channel characteristic of HVS for image perception.To this end,I proposed an image hashing based on edge detection and wavelet statistical features.This hashing converts input image to a normalized image by preprocessing,and divides the normalized image into non-overlapping blocks.For each block,multi-level wavelet decomposition is applied to obtain different sub-bands.Since the changes of different sub-bands have different effects on HVS,different sub-bands have different weights for calculation.Finally,the weighted sums of wavelet statistical features in different sub-bands are used to construct image hash.Similarity of image hash is measured by the Euclidean distance.ROC curve comparisons show that our classification performance between robustness and discrimination is significantly better than those of other algorithms.Our application in reduced reference image quality assessment is discussed,where the LIVE image database provided by the image and video engineering laboratory of the TEXAS University is used as the dataset of distorted images.The nonlinear curve fitting results show that the objective scores of the proposed algorithm are highly correlated with the different mean opinion scores(DMOS)provided by the LIVE image database.