![[ Visit Client Website ]](images/banner.gif)
The most widely held theory of facial beauty is that faces which more closely represent the average are the most beautiful. Our study sought to determine whether the theory of averageness could be demonstrated in women seeking rhinoplasty using state of the art machine learning algorithms.
METHODS
Photographic analysis consisted of 1192 pre- and post- rhinoplasty photos of women as well as 139 photos of actresses, all of whom are listed as the most beautiful women of all time per IMDB. All photos are frontal shots with the face in a neutral pose. Using a pre-trained deep convolutional network algorithm, the photos were embedded with 128 vectors for clustering analysis. Phenotyping analysis of the pre-rhinoplasty photos was conducted via parameterized Gaussian mixture models optimized via Bayesian Information Criteria (BIC) for expectation-maximization. Furthermore, facial averages were generated via a Delaunay triangulation using the 68 landmarks and facial similarity scores were computed via similarity score of two faces by computing the squared L2 distance between their representations.
RESULTS
The optimal number of pheno-groups determined by BIC, bounded by 1-5. The model assigned 410 photos to “pheno-group 1” and 782 photos to “pheno-group 2.” Beautiful actresses were more likely to be in phenotype 2 as compared to pre-rhinoplasty women (82% vs. 65%, p = .0001). Further, post-rhinoplasty women switched from phenotype 1 to phenotype 2 considerably more than they switch from 2 to 1 (21% vs. 7%, p = <.000001). Post-rhinoplasty composite faces were more similar to the “beautiful actresses” composite than the pre-rhinoplasty photos (L2 norm = 0.518 vs. 0.621).
CONCLUSION
We demonstrate that women do not become more “average” after rhinoplasty, but rather trend towards the phenotype occupied by above average beautiful women.