Feature

Characterising Caricatures

What is it that enables us to immediately recognise a caricature in a satirical cartoon?

Dr Gillian Rhodes

In the aftermath of the Watergate scandal, as Nixon's popularity plummeted, his nose and jowls grew to impossible proportions in published caricatures, yet he remained instantly recognisable. A television programme featuring caricatured puppets provoked one viewer to exclaim, "Hell, the one of Bill Cosby looks more like Bill Cosby than Bill Cosby does!"

These anecdotes illustrate the power of caricatures to capture a likeness -- their paradoxical quality of being more like the face than the face itself. Caricatures offer a challenge to cognitive psychologists who seek to understand how we recognise things. How can something so obviously distorted be so recognisable?

Susan Carey, of the Massachusetts Institute of Technology, and I have been studying caricature recognition as part of our research on how people recognise faces. We are interested in how people can recognise hundreds or even thousands of faces when they are all so similar (to see just how similar faces are, try looking at your old school photographs upside-down).

How Do We Do It?

This impressive performance poses a problem for current accounts of object recognition, which claim that objects are recognised from their component parts and the arrangement of those parts. Clearly such a scheme will not work for homogeneous objects, like faces, that have the same basic parts in the same basic arrangement.

A part-analysis could tell us that we are looking at a face, but to know whose face it is we need some way of coding the subtle differences in the shared configurations that characterise individual faces. Our research with caricatures suggests how this may be done.

We produced our images using a computerised caricature generator, created by Susan Brennan, an amateur caricaturist and a student at the MIT Media Lab.

Brennan decided to automate the steps she used to create a caricature. Her programme makes a caricature in three steps: First, a photograph of the "victim" is digitised and fixed landmark points are located on the face. These points are marked using a mouse and then the programme "joins the dots" to produce a simple line drawing image of the face.

Second, the drawing is compared with that of a norm or average face and the programme identifies corresponding points on the two images, such as the tip of the nose or the peak of the eyebrow. Different norms can be used for structurally distinct classes of face, such as males and females or young and old faces.

In the third and final step, the programme produces a caricature by exaggerating all the differences between the corresponding pairs of points by an amount specified by the user, say 50%. "Anticaricatures" are created by moving the points on the victim's face closer to the corresponding points on the norm.

Using these computer-generated images, we have confirmed that caricatures, which exaggerate distinctive information, can be identified at least as well as undistorted faces. Some are even recognised better than the undistorted images -- making them, in effect, "superportraits".

In contrast, anticaricatures, which reduce distinctive information, are very difficult to recognise. Similar results have been obtained by researchers at St Andrews University using photographic quality caricatures.

Important Deviations

The effectiveness of caricatures, which exaggerate how each face differs from a norm or average face, highlights the importance of how a face deviates from the norm -- indeed something may only become a crucial feature for recognition when it differs from the average.

The power of caricatures also suggests that we code, and hence remember, subtle variations by comparing the target face with a norm and noting how it deviates from it.

Further support for this norm-based coding idea comes from our finding that faces become quite unrecognisable if the norm-deviation features are disrupted. Instead of moving away from or towards the norm to produce caricatures and anticaricatures, we distorted the facial images by moving the points orthogonally (at right angles) to the norm-deviation direction.

Different From the Norm

If we code distinctive features as deviations from a norm, then these "lateral" caricatures should be very difficult to recognise, as indeed they are. When we compared performance on 50% caricatures, undistorted images, 50% anticaricatures and 50% laterals, performance was worst on the laterals, which were almost impossible to recognise.

Our work with caricatures suggests that we code the distinctive features of faces as deviations from a spatial norm or average face. This kind of norm-based coding is a clever solution to the homogeneity problem presented by faces because it exploits the very homogeneity that creates a problem for a part-based recognition system.

It is because faces have the same parts in the same basic arrangement that they can be averaged (essentially superimposed) to give the ideal image against which to assess each individual's unique distinguishing features.

Faces provide the most dramatic illustration of the homogeneity problem, but they are not the only homogeneous objects we must distinguish. Many biologically significant discriminations between different kinds of animals and plants, as well as more mundane discriminations between different cars or chairs, rely to some degree on an ability to code variations in a shared configuration.

Currently we are investigating whether norm-based coding is a general-purpose method of coding variations or whether it is part of a special face-recognition system.

Our main goal is to understand how we mentally represent and recognise visual objects. However, our research also has some interesting potential applications. For example, it suggests that caricatures might be useful aids for recalling faces in forensic settings and computer face recognition systems might benefit from utilising this coding.

Dr Rhodes is a senior lecturer in psychology at Canterbury University.