Implementation and benchmarking of perceptual image hash functions

Abstract

Perceptual image hash functions produce hash values based on the image's visual appearance. A perceptual hash can also be referred to as e.g. a robust hash or a ngerprint. Such a function calculates similar hash values for similar images, whereas for dissimilar images dissimilar hash values are calculated. Finally, using an adequate distance or similarity function to compare two perceptual hash values, it can be decided whether two images are perceptually di erent or not. Perceptual image hash functions can be used e.g. for the identi cation or integrity veri cation of images. This thesis proposes a novel benchmarking framework, called Rihamark, for perceptual image hash functions. Subsequently, four di erent percep- tual image hash functions were benchmarked: A discrete Cosine transform (DCT) based , a Marr-Hildreth operator based, a radial variance based and a block mean value based image hash function. pHash, an open source im- plementation of various perceptual hash functions, was used to benchmark the rst three functions. The latter, the block mean value based image hash function was implemented by the author of this thesis himself. The block mean value based image hash function outperforms the other hash functions in terms of speed. The DCT based image hash function is the slowest. Although the Marr-Hildreth operator based image hash function is not the fastest nor the most robust, it o ers by far the best discriminiative abilities. Interestingly enough, the performance in terms of discriminiative ability does not depend on the content of the images. That is, no matter whether the visual appearance of the images compared was very similar or not, the performance of the particular hash function did not change sig- ni cantly. Di erent image operations, like horizontal ipping, rotating or resizing, were used to test the robustness of the image hash functions. An interesting result is that none of the tested image hash function is robust against ipping an image horizontally.