J Forensic Sci, July 2010, Vol. 55, No. 4 doi: 10.1111/j.1556-4029.2010.01425.x Available online at: interscience.wiley.com
TECHNICAL NOTE PHYSICAL ANTHROPOLOGY
Xanth D. G. Mallett,1 Ph.D.; Ian Dryden,2 Ph.D.; Richard Vorder Bruegge,3 Ph.D.; and Martin Evison,4 Ph.D.
An Exploration of Sample Representativeness in Anthropometric Facial Comparison*
ABSTRACT: Faces are assumed to be unique, but their use in court has remained problematic as no method of comparison with known error rates has been accepted by the scientific community. Rather than relying on the assumed uniqueness of facial features, previous research has been directed at estimations of face shape frequency. Here, the influence of age, sex, and ancestry on variation was investigated. Statistical shape analysis was used to examine the necessity for sub-divisions in forensic comparisons, using a large sample of facial images on which 30 anthropometric landmark points had been placed in 3D. Results showed a clear pattern of separation of the sexes in all age groups, and in different age groups in men. It was concluded that sub-division of databases by sex will be necessary in forensic comparisons. Sub-division by age may be necessary in men (although not necessarily in women), and may be necessary by ancestry.
KEYWORDS: forensic science, facial identification, biometrics, anthropometry, principal components analysis, shape analysis It is a common assumption that faces are unique. They are ubiquitous and—as a potential means of human identification—surpass dermatoglyphic fingerprints, the dentition, and DNA in being remotely detectable. To be of value to the courts, however, a scientifically defensible approach to facial comparison is required which— rather than relying on an assumption of discernable uniqueness— will offer estimates of face shape frequency derived from large database studies (1–3). In previous research (4), the Magna database (5) was used to identify key components of normal face shape variation. In this study, the same database was used to investigate the influence of age, sex, and ancestry on variation. The representativeness of sub-categories of age and sex was examined in further detail, and the necessity for sub-division of databases in forensic facial comparisons was explored. Materials and Methods The Magna database (5) consists of facial images collected from healthy volunteers using a digital stereophotogrammetric system (Geometrix FaceVision 802 Series Biometric Camera; ALIVE Tech, Cumming, GA). Volunteers were classified by age and sex, and ancestry as self-declared according to the United Kingdom Census categories (see Tables 1 and 2). A software tool (Forensic Analyzer v.1.3; ALIVE Tech) was used to place up to 30 three1
Centre for Anatomy and Human Identification, University of Dundee, Dundee DD1 5EH, U.K. 2 Department of Statistics, University of South Carolina, Columbia, SC 29208. 3 Forensic Video Audio and Image Analysis Unit, Federal Bureau of Investigation, Building 27958A, Quantico, VA 22135. 4 Forensic Science Program, University of Toronto, Mississauga, ON L5L 1C6, Canada. *Funding provided by Technical Support Working Group (T216E). Received 30 Jan. 2009; and in revised form 16 April 2009; accepted 25 April 2009.
2010 American Academy of Forensic Sciences
dimensional (3D) anthropometric landmarks on each facial image. The landmarks selected were those demonstrating maximum intersubject variation relative to intra-subject variation (see Fig. 1 and Table 3), hence offering the greatest potential for the statistical evaluation of variation in size and shape. Each facial image was landmarked twice by different observers. Only landmark sets with all 30 landmarks present were included in this study (960 women and 2294 men, from 1968 individual subjects). Patterns of variation in the sample were analyzed using statistical shape analysis (7). Procrustes registration was used to remove differences because of position and rotation. As size differences between subject faces represent real variation, Procrustes registration was undertaken without scaling. Principal components (PC) analysis was then used to decompose differences because of variation in size and shape in different sub-groups in the sample. Statistical programming was undertaken in R (8). Results and Discussion Plots of the first three PC scores underlying size and shape variation in men and women in 15–34 (Fig. 2) and 35–54 (Fig. 3) years of age ranges confirm earlier observations (4) of sexual dimorphism, and of age-related trends in face size and shape variation that affect men and women differently. To assess whether the sample as a whole could be considered representative of the sub-categories of sex and age within it, a closer examination was made of age group categories. Individuals under 16 years of age were excluded to partly mitigate confounding effects of rapid changes in facial morphology affecting juveniles during development. Analysis of the first two PC scores underlying variation by age group and sex (Fig. 4) in this sample again shows a clear pattern of separation of the sexes in all age groups. Separation increases with age and is underlain by the first PC score, which is attributed 1025
1026
JOURNAL OF FORENSIC SCIENCES
TABLE 1—Distribution of volunteers in the Magna database by sex and age. Age group 14–19 20–24 25–29 30–34 35–39 40–44 45–49 50–54 55–59 60–64 65+ Total
Females
Males
Label
194 71 98 173 291 247 127 53 55 44 41 1394
188 68 103 188 325 361 181 111 65 54 77 1721
al c’
TABLE 2—Distribution of volunteers in the Magna database by sex and ancestry. Census category White British Other white background White and black Caribbean White and black African White and Asian Other mixed background Indian Pakistani Any other Asian background Caribbean African Other black background Chinese Any other Total
TABLE 3—Description of the 30 landmarks used in the study (and shown in Fig. 1).
Females
Males
1265 56 4 1 4 6 14 4 6 2 7 3 18 4 1394
1553 78 3 2 6 3 17 11 11 8 7 1 17 4 1721
FIG. 1—Illustration of the 30 landmark sites collected from volunteers using the Geometrix FaceVision FV802 Series Biometric Camera. Bilateral landmarks are in green.
Name
ch
Alar (l, r) Highest point of columella (l, r) Cheilion (l, r)
en
Endocanthion (l, r)
ex
Exocanthion (l, r)
g
Glabella
li ls obi p
Labiale inferius Labiale superius Otobasion inferius (l, r) Pupil (l, r)
pa
Postaurale (l, r)
pg pi prn sa
Pogonion Palpebrale inferius (l, r) Pronasale Superaurale (l, r)
sba
Subaurale (l, r)
se
Sellion
sl
Sublabiale
sto
Stomion
Description—after Farkas 1994 (6) The most lateral point on each alar contour The point on each columella crest, level with the tip of the corresponding nostril The point located at each labial commissure The point at the inner commissure of the eye fissure The point at the outer commissure of the eye fissure The most prominent midline point between the eyebrows The midpoint of the lower vermillion line The midpoint of the upper vermillion line The point of attachment of the ear lobe to the cheek Determined when the head is in the rest position and the eye is looking straight forward The most posterior point on the free margin of the ear The most anterior midpoint of the chin The lowest point in the mid-portion of the free margin of each lower eyelid The most protruded point of the apex nasi The highest point on the free margin of the auricle The lowest point on the free margin of the ear lobe The deepest landmark located in the bottom of the nasofrontal angle Determines the lower border of the lower lip and upper border of the chin The imaginary point at the crossing of the vertical facial midline and the horizontal labial fissure between gently closed lips, with the teeth shut in the natural position
to size (4). This size effect is also notable in men considered alone (Fig. 5), where the older age groups have a higher first PC score than the youngest group here. In women (Fig. 6), there is a general tendency for the first PC score to be higher for the older age groups here, but the trend is not remarkable and no clear differentiation between the youngest age group and the older groups is evident. It is tentatively anticipated that male faces in the sample complete the development to the adult form in the younger age group and that further changes as a result of aging are relatively subtle, whereas female faces retain a youthful morphology in adulthood. To explore the effect of age sub-division in men and women separately, means and standard deviations of the first four PC scores were compared by 3-year age ranges for each sex (see Tables 4 and 5). Results of the PC analyses (Figs. 2–7) indicate that sub-division of men in the database may be required for use in forensic facial comparison, although the implication is less evident for women. Examination of the means and standard deviations of the first four PC scores underlying face size and shape variation (Tables 4 and 5) may allow ages to be identified at which sub-divisions are most appropriate. For example, means and standard deviations in the first PC scores in men appear to stabilize after 30 years of age. The means and standard deviations in the next three PC scores that underlie size and shape variation do not become comparable at 30 years, however, and first PC score is attributed primarily to size. The significance of size in facial comparison is complicated in that
MALLETT ET AL. • ANTHROPOMETRIC FACIAL COMPARISON
1027
FIG. 2—Principal components analysis of size and shape variation in males and females for age groups in lower age ranges. Males 15–24: black; females 15–24: red; males 25–34: green; females 25–34: blue.
FIG. 3—Principal components analysis of size and shape variation in males and females for selected age groups in upper age ranges. Males 35–44: black; females 35–44: red; males 45–54: green; females 45–54: blue.
most contemporary questioned facial images are not of known size. By implication, in the event of a proposed match, estimation of the frequency of the given face shape must be independent of size. In the PC analysis of men and women aged 35–54 (Fig. 3), two landmark datasets representing independent landmarkings of a single male individual were detected far from the main cluster of men and even on the edge of the female distribution (seen to the left of Fig. 3). This extreme outlier does not appear to be in error, but rather a self-declared male volunteer who may be a transgendered or intersex individual. Finally, a PC analysis of size and shape variation attributable to ancestry (Fig. 7) reveals some trends affecting the United Kingdom
Census ethnic categories relative to each other, but not relative to the White British component, which predominates numerically in the sample (see Tables 2 and 6). Conclusions These results confirm that sex, age, and ancestry affect variation in face size and shape. PC analysis indicates that sub-division of databases by sex will be necessary in forensic facial comparison. Sub-division according to age group may be necessary in men (although not necessarily in women) and may be necessary according to ancestry—although a bias toward older age groups in the
1028
JOURNAL OF FORENSIC SCIENCES
FIG. 4—Pair wise plots of the first two principal components scores underlying face size and shape variation for males (black circle) and females (red triangle), in four different age groups.
FIG. 5—Pair wise plots of the first four principal components scores underlying face size and shape variation in males for four different age groups: 15– 24 years: black circle; 25–34 years: red triangle; 35–44 years: green ‘+’; 45–54 years: blue ‘·’; over 55 years: cyan diamond.
MALLETT ET AL. • ANTHROPOMETRIC FACIAL COMPARISON
1029
FIG. 6—Pair wise plots of the first four principal components scores underlying face size and shape variation in females for four different age groups: 15–24 years: black circle; 25–34 years: red triangle; 35–44 years: green ‘+’; 45–54 years: blue ‘·’; over 55 years: cyan diamond.
FIG. 7—Principal components analysis of overall size and shape variation classified into five ancestry groups. White: small cyan diamond; Mixed Ancestry: black circle; Asian: red triangle; Black: green ‘+’; Chinese or other: blue ‘·.’
1030
JOURNAL OF FORENSIC SCIENCES TABLE 4—Means and standard deviations (SD) for the first four principal components (PC) scores in men. PC1
Age 16–18 19–21 22–24 25–27 28–30 31–33 34–36 37–39 40–42
PC2
PC3
PC4
n
Mean
SD
Mean
SD
Mean
SD
Mean
SD
63 69 69 100 107 160 277 292 316
)5.7 2.8 2.6 3.2 6.5 11.7 10.4 11.0 11.6
13.2 11.0 12.8 12.5 10.5 12.7 11.5 12.3 12.6
)2.9 )1.9 )5.2 1.9 0.2 )1.4 )0.3 0.5 0.6
8.2 10.6 12.0 11.5 12.2 12.0 10.3 9.2 10.4
)1.9 )4.0 )3.3 )1.4 )5.0 )2.8 )0.4 0.7 0.2
7.7 9.0 8.4 9.5 11.1 9.9 9.6 8.7 9.0
)6.4 )6.3 )2.6 )4.8 )1.6 )0.5 0.1 0.0 0.2
5.5 7.5 7.2 7.7 8.0 7.2 7.6 7.0 7.5
TABLE 5—Means and standard deviations (SD) for the first four principal components (PC) scores in women. PC1 Age 16–18 19–21 22–24 25–27 28–30 31–33 34–36 37–39 40–42
PC2
PC3
PC4
n
Mean
SD
Mean
SD
Mean
SD
Mean
SD
49 49 41 58 53 70 127 111 120
)25.6 )21.6 )23.0 )21.9 )23.1 )19.3 )21.5 )23.2 )22.7
9.2 13.2 9.1 11.7 11.9 12.1 12.6 10.4 11.4
)0.2 )0.1 )2.4 )2.5 )5.4 )4.1 )4.5 )3.4 )3.0
8.7 9.7 9.9 10.8 8.3 8.3 8.2 8.4 10.9
)1.7 )4.0 )1.8 )4.5 0.6 )1.4 )0.2 0.0 0.8
9.6 8.2 8.9 8.7 9.1 7.7 7.8 8.5 9.3
)1.3 )3.7 )1.3 )3.1 )1.4 1.0 0.2 1.3 1.0
5.3 5.7 8.3 6.6 5.8 5.5 6.8 6.8 5.8
TABLE 6—Means and standard deviations (SD) for the first four principal components (PC) scores by ancestry. PC1
PC2
PC3
PC4
Ancestry
n
Mean
SD
Mean
SD
Mean
SD
Mean
SD
White Mixed Asian Black Chinese
3096 27 64 30 37
0.2 )2.7 )7.3 7.5 )5.3
19.0 14.3 16.3 20.0 21.2
0.0 )6.3 0.6 )5.6 6.7
10.3 8.7 10.3 15.4 11.6
0.2 )0.8 )5.7 )7.3 )1.9
9.2 8.4 10.0 12.0 7.9
0.2 )3.8 )4.1 )8.5 2.5
7.5 6.3 6.8 7.6 8.5
Magna database (see Table 1) means that sub-samples in the younger age groups were small in this analysis. Persons of non-White ancestry were similarly under-represented (see Table 2). Caution may be necessary in sub-dividing databases according to sex, however, in view of the complexities of chosen categories of gender and intersex conditions. One potential source of bias was also identified. Of the 3024 individuals in the sample, only datasets landmarked at all 30 possible points—originating from 1968 individuals—were subjected to PC analysis. Commonly excluded cases include women in which the landmarks of the ear are covered by hair, and men in which the landmarks of the mouth are obscured by beards and mustaches. Further research using the Magna database will permit the elaboration of prototypic tools for anthropometric matching or exclusion of faces and face shape frequency estimation (9). Research on the Magna database may also permit the incorporation of anthroscopic features such as scars, moles, blemishes, and tattoos—as well as shape features between landmarks—into those tools. Further data collection of facial images from volunteers of nonWhite ancestry and from younger individuals may be necessary to ultimately establish database sub-divisions required for forensic facial comparison.
Acknowledgments The authors acknowledge the contribution of the research assistants and public volunteers at the Magna Science Adventure Centre, Rotherham, U.K. The research proposal was reviewed by the SchHARR ethics committee of the University of Sheffield, U.K. References 1. Saks MJ, Koehler JJ. The coming paradigm shift in forensic identification science. Nature 2005;309:892–5. 2. Mardia KV, Coombes A, Kirkbride J, Linney A, Bowie LJ. On statistical problems with face identification from photographs. J Appl Stat 1996;23(6):655–76. 3. Evison MP. Anthropometry of the face: a review of the traditional methods of craniofacial measurement and their application to the anthropometry of photographic images. Proceedings of the Third UK National Conference on Craniofacial Identification; 2000 May; Manchester (UK). Manchester (UK): University of Manchester Department of Art in Medicine, 2000;7. 4. Evison MP, Dryden IL, Fieller NRJ, Mallett XGD, Morecroft L, Schofield D, et al. Key parameters of face shape variation in 3D in a large sample. J Forensic Sci 2010;55(1):159–62.
MALLETT ET AL. • ANTHROPOMETRIC FACIAL COMPARISON 5. Evison MP, Vorder Bruegge RW. The Magna Database: a database of three-dimensional facial images for research in human identification and recognition. Forensic Sci Commun 2008;10(2), [Online], http://www. fbi.gov/hq/lab/fsc/backissu/april2008/research/2008_04_research01.htm (accessed April 16, 2009). 6. Farkas LG. Anthropometry of the head and face, 2nd edn. New York, NY: Raven Press, 1994. 7. Dryden IL, Mardia KV. Statistical shape analysis. Chichester, UK: Wiley, 1998. 8. R Development Core Team. R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing, 2009, http://www.r-project.org/ (accessed April 16, 2009).
1031
9. Evison MP, Vorder Bruegge RW. Computer-aided forensic facial comparison: scientific and technical aspects. New York, NY: CRC Press ⁄ Taylor and Francis, 2010. Additional information and reprint requests: Martin Paul Evison, Ph.D. Forensic Science Program University of Toronto Mississauga 3359 Mississauga Road North Mississauga, ON L5L 1C6 Canada E-mail:
[email protected]
