An investigation into the use of subspace methods for face detection

In this work, we investigate the use of subspace methods as a representation for the human face-space and how to apply them to face detection for low resolution images (19 × 19 pixel images). We compare between different subspace paradigms, namely, principal component analysis (PCA), linear discriminant analysis (LDA) and kernel linear discriminant analysis (KLDA). We find that particularly the eigenface corresponding to the smallest non-zero eigenvalue is useful in detecting non-face images as outliers. We also find that using this eigenface in conjunction with the basis computed by LDA gives better results in comparison with kernel LDA when tested on a very large test-set of 36,806 images and with much lower computation required. Furthermore, we compare the computational complexity of our method with Rowley's face detector [1], which is considered as the most robust real-time face detector [2].