Face Phylogeny Tree: Deducing Relationships Between Near-Duplicate Face Images Using Legendre Polynomials and Radial Basis Functions

Photometric transformations such as brightness and contrast adjustment are widely used to enhance the visual aesthetics of face images. A face image can be repeatedly modified using such transformations, resulting in a set of near duplicate images. In this work, we tackle the difficult problem of generating an Image Phylogeny Tree (IPT) from near-duplicate face images. An IPT captures the structure of evolution between images. We first assemble a dataset of near-duplicate face images using random sequences of photometric transformations. Secondly, we develop an algorithm to model the pairwise transformation using Legendre polynomials and Gaussian radial basis functions. Given a set of near-duplicate face images, we utilize these basis functions to model the pairwise relationship between these images. Thirdly, we develop a likelihood-ratio based scheme to distinguish between forward transformations and reverse transformations. The likelihood-ratio value itself is used to develop an asymmetric similarity matrix that depicts the relationship between pairs of images. Finally, we deduce the IPT from this matrix. Experiments conducted on images from the LFW dataset involving 2,727 IPTs convey the promise of our method to confront this challenging problem in image forensics, exhibiting 80% root node identification accuracy for near-duplicate face images.