A probabilistic measure of similarity for latent fingerprints
In forensics a latent print is a mark left on a surface by a human friction ridge pattern such as a fingerprint. The task of latent print examination involves comparing a latent print to a known (or exemplar) print. Latent print examiners make a hard decision of identification/exclusion/no conclusion. Although latent prints have served as crucial evidence in forensics, there have been high-profile cases highlighting the fact that human/machine errors can occur. Errors arise due to the latent print being of poor quality, small area, large nonlinear distortion and unreliable extracted features. The goal of this research is to develop a statistical model for the task of comparing a latent print to a known print so that final opinion can be expressed probabilistically and the accuracy of the decision can be improved. A likelihood ratio (LR) measure is commonly used to quantify strength of evidence.It is the ratio between the joint probability that the evidence and known come from the same source, and the joint probability that the two come from different sources. Since the joint probability over a large number of correlated variables is computationally intractable it is often replaced by the distribution over a discriminative similarity measure, e.g., a matching score, thereby transforming a multivariate problem into a univariate one. However there is a severe loss of information in such a transformation . It is not possible for such a deterministic quantity to capture inherent uncertainty about the latent fingerprint or combine with other results. Instead of a deterministic scalar measure of similarity we obtain a joint probabilistic distribution about the similarity and furtherly infer a probability from maximum a posteriori(MAP) estimate. The similarly between two fingerprints is measured as a probability of establishing most likely correspondences between two minutiae sets via maximizing the posterior of associating them. We transform the fingerprint correspondence problem into an optimal mapping between two sets of features and formulate it in a Bayesian framework using a Markov Random Field (MRF). In this framework, the task is to find the maximum a posteriori (MAP) estimate of associating the two sets of minutiae. A multivariate MRF representing the mapping between known and latent prints defines the prior distribution. Then we incorporate observation potentials into the potentials of MRF by defining a Gaussian likelihood function based on observed dissimilarities between the latent print and the known. Since the number of variables and the model structure of MRF is different in each comparison, a generative model, learnt from a large sized training data set, is built to estimate node/edge parameters of the MRF. This model takes into account distortions of corresponding minutiae and provides prior information about the mapping. The discriminative power of the method is evaluated and compared with baseline methods and results have demonstrated significant improvement of our method.