Suppose, you want to "mix" two pictures of human faces in a specific way, so that the resulting compound image shares the properties of both original faces, i.e. it looks like an "intermediate" between the two originals. How would you do this? The technique used in order to generate compound images is referred to as "morphing", and in the following sections, you will learn a great deal about how this works, and how we used this technique for the purposes of our study.
Before we could start with morphing, we took standardized digital photographs of 64 female and 32 male faces aged 17-29 years, including eight photo models. In a preliminary test, these faces were randomly presented to test subjects using a self-programmed presentation Software (Authorware 5.0 package). Test subjects rated the attractiveness of the faces on a seven-point Likert scale ranging from 1 (very unattractive) to 7 (very attractive). On the basis of these ratings, the pictures of the faces were ranked for further use according to their average attractiveness values.
Using a morphing software (Morpher 3.0, freeware), new 50:50 percent compound images were generated from each two original faces using a binary tree scheme and following the order of the attractiveness values for the faces. For example, a compound image (w1-2) was generated by combining the least attractive female face (w1) with the second-least attractive female face (w2); in a next step, w3 and w4 were combined to w3-4, and so on. The next generation of pictures was then produced by combining w1-2 and w3-4 to w1-4, then w1-4 with w5-8 to w1-8 and so on, resulting in two single compound images w1-64 for the 64 female and m1-32 for the 32 male faces.
If you mix the left picture with the one in the middle, the result is an intermediate between the two faces (see right picture).
But how does morphing work? Special morphing algorithms are used, and corresponding reference points need to be defined for each of the original images (for example, the tip of the nose in the left and right picture). The reference points are then connected by reference lines, so that areas of equal texture and form are enclosed - for example, the mouth of each face is surrounded by a completely closed reference line (see left figure). To calculate the intermediate face, the coordinates of the reference points and the colour values for each pixel are averaged.
In our study, we had to define more than 500 reference points for each morphed image (i.e. 250 in each of the original faces). If you sum everything up, we had to specify over 75.000 reference points in total. In this way, we were able to produce pictures of average faces that are almost indistinguishable from 'normal' everyday faces, making the faces created by morphing easily comparable to the original photographs taken. Thus, the applicability of our results is likely to be higher than in previous studies, in which lower-resolution compound images have been produced.
left: image of an average face (made from 16 female faces) that has been created using the 'overlay technique' as it has been widely used in former studies on human facial attractiveness (from Grammer, 1995, "Signale der Liebe", dtv Verlag, Germany). It was created in 1990 with state-of-the-art technology of that time. Nowadays, with more powerful processors and improved software programs, it is possible to do research with much more realistic images, like the one shown to the right, which has been calculated using the 'morphing' technique from 16 original faces, likewise.
Again (in analogy to our preliminary attractiveness rating described in the summary) we asked test subjects to rate the attractiveness of the new intermediate images on a 7-point Likert scale. In this way, we were able to compare the attractiveness values for the original faces and the morphs.
But what's so interesting about this? Well, in psychology, the so-called
"averageness is attractiveness" hypothesis has become very popular, stating
that faces resembling the average of a population are regarded most attractive
(Langlois & Roggmann, 1990). We wanted to challenge this hypothesis
by using much more realistic images created by means of the morphing technique
- for most of the data supporting the averageness hypothesis have been
collected in the 1990ies, when only pictures of quite low quality were
available for research. If you're interested in what we found out, proceed
to the section "average
Last modified: 10-01-2012, webmaster