ReFace (Reality Enhancement Facial Approximation by Computational Estimation) is a prototype facial approximation software program developed by the Federal Bureau of Investigation (FBI) in conjunction with GE Global Research. The prototype extrapolates an “approximation” of a face from a skull using a database of computed tomography (CT) scans of living individuals. The test set consisted of CT scans of 53 articulated human skulls from the William M. Bass Donated Skeletal Collection and the William M. Bass Forensic Skeletal Collection, which are curated at the University of Tennessee in Knoxville.
Through the Federal Bureau of Investigation’s Visiting Scientist Program, an educational opportunity administered by the Oak Ridge Institute of Science and Education (ORISE), the researcher conducted an independent validation of this software in two phases. Phase 1 tested and evaluated the software performance, resulting in improvements to the software and the development of standardized protocol for articulation, packaging, and preparation of human skulls for CT scans. Phase 2 validated the accuracy of the software in the production of facial approximations from human skulls using face pools and resemblance ratings.
In Phase 2, computerized facial approximations were visually compared with antemortem photographs by four participant groups (N = 103). Ten test subjects of European ancestry (six females and four males) were selected for a photographic validation by face pool and resemblance rating validation tests. Participants were asked to choose the face pool photograph that most closely resembled the facial approximation produced by ReFace. In the second test, the same volunteers were asked to rate (on a scale of 1 to 5) how closely ReFace facial approximations of target subjects resembled an antemortem photograph.
In the Face Pool Validation Test, nine out of ten target subjects were correctly identified above random chance, and the frequency distribution was statistically above chance expectations for nine out of ten target subjects (p < .01). The mean hit rate for all subjects was 24% (10% above random chance). There were no significant differences in the hit rates between male participants (67%) and females participants (33%), or between participant groups. All participants were non-experts. Male target subjects received higher numbers of correct responses than female target subjects. The overall ratings for the Resemblance Rating Validation Test were 13% none, 24% slight, 22% approximate, 25% close, and 16% strong. The majority of subjects were rated as close resemblance (six subjects), strong resemblance (one subject), approximate resemblance (one subject), and slight resemblance (one subject). The foil comparison received an equal number of ratings for no resemblance (30.5%) and slight resemblance (30.5%).
| Identifer | oai:union.ndltd.org:UTENN/oai:trace.tennessee.edu:utk_gradthes-1207 |
| Date | 01 August 2007 |
| Creators | Moyers, Diana Kim |
| Publisher | Trace: Tennessee Research and Creative Exchange |
| Source Sets | University of Tennessee Libraries |
| Detected Language | English |
| Type | text |
| Source | Masters Theses |
Page generated in 0.0022 seconds