We found a match
Your institution may have rights to this item. Sign in to continue.
- Title
ON DESIGNING SWIR TO VISIBLE FACE MATCHING ALGORITHMS.
- Authors
Whitelam, Cameron; Bourlai, Thirimachos
- Abstract
Recent advances in facial recognition have trended towards cross-spectrally matching visible gallery face images to probe face images captured under different wavelengths of the electromagnetic spectrum. In this article, we study the problem of matching visible images to images taken in the short-wavelength infrared (SWIR) spectrum, more specifically, the 1550-nm band. There are many benefits to using the SWIR spectrum for face recognition, including covert capturing in nighttime environments as well as imaging through certain environmental conditions such as fog and smoke. However, due to the fact that the moisture in the skin tends to absorb the 1550-nm wavelength, all subjects appear to have dark or black skin tone. Because of the stark contrast between 1550-nm and visible face images, standard face recognition protocols fail to accurately match images captured using sensors operating on different bands. While preliminary work in this area resulted in fairly good performance results, it was determined that a more sophisticated approach could be developed to further improve our original face recognition algorithm in terms of (i) accuracy, (ii) speed, and (iii) adaptability, that is, the proposed algorithm should achieve good results on a wider variety of testing scenarios (diverse face datasets). More specifically, we study the advantages and limitations of our new proposed cross-spectral matching (visible to SWIR) technique when using an extended set of challenging FR scenarios. The proposed face matching algorithm is a significant improvement when compared to the original algorithm where fused texture-based scores of a large number of photometric normalization combinations between SWIR and visible images were used to achieve satisfactory recognition performance results. Our contributions are threefold. Firstly, multiple databases are considered, which represent different difficult environments, that is, multiband face images were acquired under different lighting conditions and behind different obscurants (multiple levels of tinted glass). Secondly, we demonstrate that the use of a random selection of intensity-based normalization techniques is not necessary. This is because a random combination of such techniques does not have a significant amount of discriminatory information to accurately match one subject's face to another, yielding undesirably low face-matching scores. Thirdly, we demonstrate that a smart selection of a subset of normalization techniques not only results in obtaining more accurate face recognition performance scores, but also drastically decreases the long processing time required to produce even a single face-to-face image match score. Our design also incorporates the usage of parallel processing to further boost the time needed to perform cross-spectral matching. Finally, our experiments indicate that the level of improvement in recognition accuracy is scenario dependent.
- Subjects
HUMAN facial recognition software; BIOMETRIC identification software; COMPUTER security management; COMPUTER security software; COMPUTERS
- Publication
Intel Technology Journal, 2014, Vol 18, Issue 4, p98
- ISSN
1535-864X
- Publication type
Article