Statistical hypothesis pruning for identifying faces from infrared images

摘要

A Bayesian approach to identify faces from their IR facial images amounts to testing of discrete hypotheses in presence of nuisance variables such as pose, facial expression, and thermal state. We propose an efficient, low-level technique for hypothesis pruning, i.e. shortlisting high probability subjects from given observed image(s). (This subset can be further tested using some high-level model for eventual identification.) Hypothesis pruning is accomplished using wavelet decompositions (of the observed images) followed by analysis of lower-order statistics of the coefficients. Specifically, we filter infrared (IR) images using bandpass filters and model the marginal densities of the outputs via a parametric family that was introduced by Grenader and Srivastava [IEEE Trans. Pattern Anal. Mach. Intell. 23 (2001) 424]. IR images are compared using an L2-metric between the Marginals computed directly from the parameters. Results from experiments on IR face identification and statistical pruning are presented.