Approaches for Improving Cross-Spectrum Face Recognition
LUNCH PROVIDED AFTER TALK
Benjamin S. Riggan, PhD
Electronics Engineer
US Army Research Laboratory (ALC)
Networked Sensing and Fusion Branch
Abstract:
For covert daytime or nighttime intelligence gathering, the ability to match visible face imagery from a gallery set with thermal face imagery is high desirable. Robust image descriptors that are useful for homogeneous face recognition (e.g. visible to visible) often do not achieve acceptable recognition rates when applied cross-spectrum face recognition (e.g. thermal to visible). In this work, the fundamental approach is to exploit mutual information between corresponding sets of visible and thermal images to extract an interoperable set of features, which are used to improve the cross-spectrum recognition performance. Several contributions to support this effort include the use polarization state information, cross-spectrum neural networks and discriminative classifiers, and cross-spectrum synthesis.