Face Recognition using Multiple Constrained Mutual Subspace Method

We represent sets of face images as low-dimensional subspaces and compare them after projection onto multiple constraint subspaces. We generate these constraint subspaces through ensemble learning to extract features useful for identifying people.

A person's face can look slightly different from one photograph to the next. This study proposes improving recognition by comparing collections of face images rather than relying on a single image.

We summarize each collection using a subspace: a compact representation with a small number of axes describing differences among the images. We then measure how similar a new collection is to a collection registered for a known person. Instead of comparing them in just one way, we transform both representations in several ways designed to make differences between people easier to detect.

To learn these different comparisons, we use ensemble learning, which combines results from multiple learning processes. Specifically, we use methods called Bagging and Boosting. Experiments showed that the resulting Multiple Constrained Mutual Subspace Method was effective in improving face recognition accuracy.

• Masashi Nishiyama, Osamu Yamaguchi, and Kazuhiro Fukui,
  Face Recognition with the Multiple Constrained Mutual Subspace Method, Proceedings of 5th International Conference on Audio- and Video-based Biometric Person Authentication 2005 (AVBPA), pp. 71 - 80, July 2005.
• Tomokazu Kawahara, Masashi Nishiyama, Tatsuo Kozakaya, and Osamu Yamaguchi,
  Face Recognition based on Whitening Transformation of Distribution of Subspaces, Workshop on ACCV2007 Subspace, pp. 97 - 103, November 2007.

• 西山 正志, 山口 修, 福井 和広,
  多重制約相互部分空間法を用いた顔画像認識,
  電子情報通信学会論文誌 D-II, Vol. J88-D-II, No. 8, pp. 1339 - 1348, 2005.