Estimation Of Visible Spectrum Faces From Polarimetric Thermal Faces

A method for synthesizing visible spectrum face imagery from polarimefric-thermal face imagery is presented. This work extends recent within-spectrum (i.e., visible-to-visible) reconstruction techniques for image representation understanding using convolutional neural networks. Despite the challenging task, we effectively demonstrate the ability to produce a visible image from a probe polarimetric-thermal image. Moreover, we are able to demonstrate the same capability with conventional thermal imagery, but we report a significant improvement by incorporating polarization-state information. These reconstructions, or estimates, can be used to aid human examiners performing one-to-one verification of matches retrieved from automated cross-spectrum face recognition algorithms.