This paper describes the acquisition process and content of a multispectral face database, which can be used to research on face recognition methods dealing with two of the most challenging problems in this area, i.e. partial occlusion and pose variations. Four cameras were synchronized and arranged to simultaneously capture images from visible, thermal, ultraviolet and near-infrared spectra, which had reported promising results for recognizing faces, individually. In order to simulate pose variations, each subject was asked to look forward, up, down, and to the sides, varying the point of view angle. On the other hand, partial occlusion was generated using sunglasses and a paper sheet. Additionally, three lighting changes were also included (halogen, natural and infrared). A total of 306 images were acquired by subject and 31 subjects were recruited. So, the whole database is composed of 9486 images, which are now available to other researchers. Preliminary results showed that spectra variations affect the performance of a deep learning recognition approach. As far as we know, this is the first database of faces including images from those spectra and the other variations simultaneously.
CITATION STYLE
Fonnegra, R. D., Molina, A., Pérez-Zapata, A. F., & Díaz, G. M. (2018). MSpecFace: A dataset for facial recognition in the visible, ultra violet and infrared spectra. In Communications in Computer and Information Science (Vol. 798, pp. 160–170). Springer Verlag. https://doi.org/10.1007/978-3-319-72727-1_12
Mendeley helps you to discover research relevant for your work.