A comparative study of thermal face recognition methods in unconstrained environments

摘要

The recognition of faces in unconstrained environments is a challenging problem. The aim of this work is to carry out a comparative study of face recognition methods working in the thermal spectrum (8–12 μm) that are suitable for working properly in these environments. The analyzed methods were selected by considering their performance in former comparative studies, in addition to being real-time, to requiring just one image per person, and to being fully online (no requirements of offline enrollment). Thus, in this study three local-matching methods based on histograms of Local Binary Pattern (LBP) features, on histograms of Weber Linear Descriptors (WLD), and on Gabor Jet Descriptors (GJD), as well as two global image-matching method based on Scale-Invariant Feature Transform (SIFT) Descriptors, and Speeded Up Robust Features (SURF) Descriptors, are analyzed. The methods are compared using the Equinox and UCHThermalFace databases. The use of these databases allows evaluating the methods in real-world conditions that include natural variations in illumination, indoor/outdoor setup, facial expression, pose, accessories, occlusions, and background. The UCHThermalFace database is described for the first time in this article and WLD is used for the first time in face recognition. The results of this comparative study are intended to be a guide for developers of face recognition systems. The main conclusions of this study are: (i) all analyzed methods perform very well under the conditions in which they were evaluated, except for the case of GJD that has low performance in outdoor setups; (ii) the best tradeoff between high recognition rate and fast processing speed is obtained by WLD-based methods, although the highest recognition rate in all cases is obtained by SIFT-based methods; and (iii) in experiments where the test images are acquired in an outdoor setup and the gallery images are acquired in an indoor setup, or vice versa, the performance of all evaluated methods is very low. As part of the future work, the use of normalization algorithms and calibration procedures in order to tackle this last issue will be analyzed.