Where’s Waldo, or better yet is he or anyone else even in the picture. For a human the task of looking at in image or even a scene and determining whether there is a human face present is a simple task. We have the ability to not only detect faces, but we can remember hundreds of them and pair them with the individuals they belong to.
Computers on the other hand, aren’t quite as observant as we are. In their study “A Hierarchical Neural Network For Human Face Detection” Juell and Marsh describe there efforts to build a system that could detect one or more faces in an image. In order to perform this task, they subdivided it into two parts: edge enhancement and training of the neural network.
Edge Enhancement:
As determined over a hundred years ago by Venetian painter Canaletto, the human eye does not see neighboring regions of brightness exactly as they appear. Instead the neurons of the eye and related regions are organized in such a manner so that changes in brightness are increased or enhanced. This can be seen in the following illusion constructed by Hermann. The intersections should appear to be brighter than the rest of the white lines, even though there are only two colors in the image, black and white.

Using a filter kernel that affected neighboring pixels Juell and Marsh enhanced all the images before sending them into the neural networks

Neural Networks:
They used four interconnected back propogation neural networks which were used for the following tasks: mouth detection, eye detection, nose detection, face detection. The first three “child” networks were trained on individual features of the face, in such a way that they were to disregard all other features. Then based on the outputs of each of these child networks the fourth, or parent network made a decision about whether a face was present. All three features, 4 if you count the eyes separately, had to be present and in a reasonable location in order for a face to be detected by the parent net.

Results:
Overall the authors of this paper were fairly successful in their efforts to build a system for face detection, though they did note that the “nose net” was not terribly useful, and two “eye nets” would have been better. One caveat, is that they prepared all the images a great deal before they inputed them into the edge enhancement and neural net phases, by resizing and cropping images. However, this paper was excellent do the simplicity of the ideas presented, and is a great introduction to neural networks and face detection.

Links:
A hierarchical neural network for human face detection
If the above link doesn’t work the article can be found in Pattern Recognition Vol 29 No. 5, which is available at the library website.
A good introduction to Neural Networks.