Research On Robust And Discriminative Image Perceptual Hashing

New challenges related to information security have been raised in front of the evergrowing digital image.Perceptual hashing is a one-way compression technology that extracts hash sequence from the visual content of an image.Through the similarity between hash sequences,perceptual hashing can flexibly meet some applications,and has received extensive attention in the field of image information security.The perceptual hashing algorithm should have the characteristics of robustness,discrimination,security,sensitivity,compactness,etc.,in which robustness and discrimination are the key to the effective application.In this paper,the robust and discriminative perceptual hashing algorithms are explored from the perspective of performance benchmark and practical application.The main contributions are as follows:(1)The perceptual hashing algorithm is proposed for natural scene images based on invariant vector distance.The main problems in natural scene images perceptual hashing are the poor classification between robustness and discrimination,as well as low sensitivity to near-duplicate images.To this end,the low-frequency coefficients in DCT domain are novelty extracted to calculate invariant vector distances,which are robustness to single content-preserving operations.In view of the high local similarity between the original image and its nearduplicate image,a strategy based on global feature is proposed to compensate the ability of overall discrimination for local invariant vector distance.The results on public datasets show that the proposed method effectively solves the above problems,and achieves high detection accuracy in the application of copy image detection.(2)The perceptual hashing algorithm is proposed for screen content images based on visual content understanding.Screen content images are different from natural scene images in the statistical characteristics and content understanding,which makes that the performance of existing algorithms on screen content image is limited.Thus,a perception mechanism based on the distribution of interest points is proposed to describe the importance of text regions.In view of the characteristics of high contrast and thin edge in text region,the statistical features derived from the maximum gradient are extracted to generate hash sequence,so that achieves good stability to different content-preserving operations.The proposed method not only improves the robustness and discrimination,but also obtains high accuracy of image visual quality assessment.(3)The perceptual hashing algorithm is proposed for multiple content-preserving operations based on structure-aware.When the hashing algorithm is tested with the single content-preserving operation dataset,there are some differences between the theoretical performance and the real application effect.Therefore,a large-scale dataset of multiple content-preserving operations is established to match the real scenes,and the stable features generation method is proposed to improve the robustness of multiple content-preserving operations.Based on the multi-channel characteristics of human visual system,multi-scale spatial sturcture features derived from the optimized edges are extracted to construct hash sequence.Experimental results show that the proposed method can significantly improve the robustness of multiple content-preserving operations and achieve high detection accuracy.