A modification of kernel-based Fisher discriminant analysis for face detection

Presents a modification of kernel-based Fisher discriminant analysis (FDA) for face detection. In a face detection problem, it is important to design a two-category classifier which can decide whether the given input sub-image is a face or not. There is a difficulty with training such two-category classifiers because the "non-face" class includes many images of different kinds of objects, and it is difficult to treat them all as a single class. Also, the dimension of the discriminant space constructed by the usual FDA is limited to one for two-category classification. To overcome these problems with the usual FDA, the discriminant criterion of the usual FDA is modified such that the covariance of the "face" class is minimized while the differences between the center of the "face" class and each training sample of the "non-face" class are maximized. By this modification, we can obtain a higher-dimensional discriminant space which is suitable for "face/non-face" classification. It is shown that the proposed method can outperform a support vector machine (SVM) by "face/non-face" classification experiments using the face images gathered from the available face databases and the many face images on the Web.