A benchmark for perceptual hashing based on human subjective identification

This study proposed a novel benchmark for evaluating the robustness and discriminabiliry properties of perceptual hashing algorithms. Firstly, two major problems neglected by traditional benchmark are analyzed thoroughly with a concrete experiment. One problem is the inconsistence between the subjective feeling and the objective perceptual distance, the other is the partiality of the performance for different attacks. And then, in order to overcome the problems, a new benchmark for perceptual hashing based on human subjective identification is proposed and the corresponding evaluation methods are presented by illustrative experiments and examples. Present benchmark methods are fairer and more comprehensive than the traditional methods.