Quantitative accuracy and 3D biometric matching of 388 statistically estimated facial approximations of live subjects. (June 2020)

• Record Type: Journal Article
• Authors: Simmons-Ehrhardt, Terrie L.; Monson, Keith L.; Flint, Toby; Saunders, Christopher P.
• Is Part Of: Forensic Imaging. Volume 21(2020)
• Publication Date: 2020-06
• Subjects: Facial approximation -- ReFace -- Craniofacial identification: Facial recognition -- Unidentified human remains -- Missing persons -- Linear discriminant analysis -- 3D modeling -- Digitization
• DOI: 10.1016/j.fri.2020.200377
• Languages: English
• ISSNs: 2666-2256

Abstract

Highlights: Accuracy of 3-D approximations was assessed by a geometric morphometric method. Approximations assessed relative to CT models for both sexes and four ancestries. Used 66 inter-landmark distances derived from twelve interior facial landmarks. 91% of known vs. approximation distances differed by ± 5.0 mm or less. Over 73% of known-approximation pairs were matched using a PCA-LDA classifier. Abstract: Objective: A geometric morphometric approach for testing the metric accuracy of a computerized facial approximation program is presented. Materials and methods: This study uses a leave-one-out procedure to generate average approximations (not adjusted for weight/age) for each individual in the ReFace reference database (n = 388), consisting of eight sex/ancestry groups (both sexes for each: African, Asian, European, Hispanic). Quantitative comparisons of corresponding three-dimensional (3D) known faces and approximations were analyzed using 66 inter-landmark distances (ILDs) derived from twelve interior facial landmarks. Results and conclusions: Average signed errors per ILD ranged from -0.67 mm to 1.92 mm, with most errors indicating smaller ILDs in approximations. For all individual ILD errors combined, 99.5% were within ± 10.0 mm, 90.8% were within ± 5.0 mm, 64.6% were within ± 2.5 mm, and 30.2% were within ± 1.0 mm. The largest errors were associated with cheilion and indicated underestimated mouth widths by an overall average of 1.92 mm (SD = 3.99 mm). A statistical matching model that modeled performance of automated facial recognition correctly identified 73–88% of known-approximation pairs as matches. The accuracy and matching results suggest that facial features are highly influenced by the underlying craniofacial skeleton regardless of weight/age variations, the matching model is predictive of automated facial recognition performance, and that ReFace approximations may be sufficiently accurate to biometrically match unidentified decedents to missing persons. Graphical abstract: Image, graphical abstract