analysis of perceptual image hash functions

Perceptual hash functions generate specific values on the basis of visual data of images that are called footprints or perceptual hashes. For similar and different images, such functions compute similar and different hashes, respectively. As a result, using the functions to compute differences or similarities between the hashes, it is possible to draw conclusion whether the relevant images are similar or not. Perceptual hash functions can be used to identify or verify image integrity.

This article analyzes existing methods of content-based retrieval by means of perceptual hash, and possibilities to use them to search for similar images. It considers the following methods to generate perceptual hashes: hash on average using low-pass filter, hash function based on histogram of colors, and hash function based on discrete cosine transformation. The article also investigates properties of hash function, which is a combination of methods based on discrete cosine transform and color histogram.

The studies were conducted in three areas: method speed, method of efficient distinguishing between images, resistance to image transformations (scaling, rotation, horizontal reflection).

All of considered algorithms have low computational complexity and, consequently, almost the same speed to calculate hashes. Hash functions based on discrete cosine transformations have shown the best ability to distinguish between the images. The color histogrambased function is the most resistant to transformation, but at the same time has the worst ability to distinguish between the images.

As a result of research the algorithm has been created, which is a combination of methods based on color histograms and discrete cosine transformation. Using this algorithm in solving the problem of search for similar images provides the most accurate and consistent results.

Further researches are supposed to expand a scope of considered algorithms to calculate the perceptual hashes.