Multispectral Imaging For Face Recognition Over Varying Illumination

This dissertation addresses the advantage of using multispectral narrow-band images over conventional broad-band images for improved face recognition under varying illumination. To verify the effectiveness of multispectral images for improving face recognition performance, three sequential procedures are taken into action: multispectral face image acquisition, image fusion for multispectral and spectral band selection to remove information redundancy.

Several efficient image fusion algorithms are proposed and conducted on spectral narrow-band face images in comparison to conventional images. Physics-based weighted fusion and illumination adjustment fusion make good use of spectral information in multispectral imaging process. The results demonstrate that fused narrow-band images outperform the conventional broad-band images under varying illuminations. In the case where multispectral images are acquired over severe changes in daylight, the fused images outperform conventional broad-band images by up to 78%. The success of fusing multispectral images lies in the fact that multispectral images can separate the illumination information from the reflectance of objects which is impossible for conventional broad-band images.

To reduce the information redundancy among multispectral images and simplify the imaging system, distance-based band selection is proposed where a quantitative evaluation metric is defined to evaluate and differentiate the performance of multispectral narrow-band images. This method is proved to be exceptionally robust to parameter changes.

Furthermore, complexity-guided distance-based band selection is proposed using model selection criterion for an automatic selection. The performance of selected bands outperforms the conventional images by up to 15%. From the significant performance improvement via distance-based band selection and complexity-guided distance-based band selection, we prove that specific facial information carried in certain narrow-band spectral images can enhance face recognition performance compared to broad-band images. In addition, both algorithms are proved to be independent to recognition engines.

Significant performance improvement is achieved by proposed image fusion and band selection algorithms under varying illumination including outdoor daylight conditions. Our proposed imaging system and image processing algorithms lead to a new avenue of automatic face recognition system towards a better recognition performance than the conventional peer system over varying illuminations.