This paper describes a method that uses a mixture of subspaces to detect faces with facial expression and facial features, faces in different poses, and faces under different lighting conditions. First, we use Kohonen's Self Organizing Map to divide the face and nonface spaces into a finite number of classes. A byproduct of the training process is a prototype of each face and noface class. Within each face and nonface class, we use Fisher Linear Discriminant to find a optimal projection matrix, thereby reducing the high dimensional samples down to lower dimensional subspaces. Gaussian density functions are then used to model the distributions of face and nonface samples in the subspaces. An input pattern is classified to be a face if its distance metrics to face clusters are less than those to nonface clusters. Experimental results show that such approach is feasible to detect faces regardless of their poses, facial expressions and lighting conditions.
Back to face detection page