Face recognition by statistical analysis of feature detectors

摘要

A successful face recognition system calculates similarity of face images based on the activation of multiscale and multiorientation Gabor kernels, but without utilizing any statistical properties of the given face data [M. Lades, J.C. Vortbrüggen, J. Buhmann, J. Lange, C. von der Malsburg, R.P. Würtz, W. Konen, Distortion invariant object recognition in the dynamic link architecture, IEEE Transactions on Computers 42 (1993) 300–311]. A method has been developed to weight the contribution of each element (1920 kernels) in the representation according to its power of predicting similarity of faces. The same statistical method has also been used to assess how changes in orientation (horizontal and vertical), expression, illumination and background contribute to the overall variance in the kernel activations. It was shown on a Caucasian and a Japanese image-set that weighting the elements in the representation according to their discriminative power would increase recognition performance. It has also been demonstrated that the weighting method is particularly useful when data compression is a key requirement. The advantages of the weighting scheme were also verified by double cross-validation.