One approach to this problem is to divide the image into many small overlapping patches at different locations, scales and orientations, and to classify each such patch based on whether it contains face-like texture or not. This is called a sliding window detector. The system then returns those locations where the probability of face is sufficiently high.