Analysis of Perceptual Hashing for Threshold Values

Perceptual hashing algorithms have been used in image similarity, image authentication and image copyright problem. There have been several attempts made in the existing literature to design perceptual hashing algorithms using manual or learning based feature extraction. Perceptual hashing algorithms have been used to check whether two images are similar or not on the basis of the distance between images. If distance between images is below the threshold value then images are similar otherwise different. The usage of a small threshold value produces false negatives whereas a large threshold value produces false positives. Thus the deciding factor for the similarity of images is threshold value, which needs to be chosen very carefully. In this paper, the problem of finding an optimal value has been considered in several perceptual hashing algorithms on three different datasets. Our results show that the optimal threshold value for the pHash, Blockhash and PDQ is 0.1 whereas optimal threshold value for Blockhash is 0.2. Further the discrimination capability of the perceptual hashing algorithms has been evaluated at different threshold values.