Face Recognition using Classified Appearance-based Quotient Image

We classify facial appearance into components such as diffuse reflection, specular reflection, and shadows, then synthesize an image with normalized illumination. Using basis images that represent diffuse reflection on a generic face improves face recognition under varying lighting conditions.

The same face can look different when the lighting direction changes. Shadows may appear, or parts of the skin may look shiny. These changes make computer-based face recognition harder. This study proposes creating a face image with reduced lighting effects so that faces are easier to recognize across different conditions.

The method builds on comparing brightness at each location with a smoothed version of the surrounding brightness. Strong reflections and shadows cause difficulties: from one image alone, it is hard to tell whether a bright or dark area comes from skin color or lighting. We therefore use a model of how light falls on a generic face. It helps distinguish broadly spread reflection, shiny reflection, and shadows before correcting the image.

We evaluated the method on Yale Face Database B and on images captured in a real-world setting, using only one training image for each person. The results showed that distinguishing these appearances and reducing lighting effects was effective for identifying faces.

• Masashi Nishiyama, Tatsuo Kozakaya, and Osamu Yamaguchi,
  Illumination Normalization using Quotient Image-based Techniques, Recent Advances in Face Recognition edited by K. Delac, M. Grgic and M. S. Bartlett, Chapter 8, pp. 97 - 108, I-Tech Education and Publishing, December 2008.
• Masashi Nishiyama, and Osamu Yamaguchi,
  Face Recognition Using the Classified Appearance-based Quotient Image, Proceedings of 7th IEEE International Conference on Automatic Face and Gesture Recognition (FG), pp. 49 - 54, April 2006.

• 西山 正志, 山口 修,
  物体表面の見え属性分類に基づく Self Quotient Image を用いた顔認識 , 信学技報 PRMU 2005-89, pp. 33 - 38, October 2005. [研究奨励賞]