Photoanthropometry in forensics: Comparison of facial images with frontal and lateral views

Photoanthropometry quantify the facial proportions of an individual facilitating the comparison of facial patterns for human identification. The coordinates and vertical distances in pixels of the photoanthropometric landmarks on images of the same individual in frontal and profile views were analyzed and compared. A total of 116 pairs of photographs of Brazilian individuals were evaluated. The photographs were adjusted in size and rotation, and marked in the software Two-dimensional Forensic Facial Analysis System. For each face, 16 landmarks were considered: glabella (g), nasion (n), ectocanthion (ec), pronasale (prn), subnasale (sn), alare (al), cheilion (ch), upper lip (ls), lower lip (li), stomion (sto), labiomental (lm), gnathion (gn), superaurale (sa), subaurale (sba), postaurale (pa) and upper ear lobe (slb); the xand y-coordinates of each landmark were obtained. Twenty-seven vertical distances between the points were proposed, which were measured by subtracting the values of the y-coordinate. The data were analyzed descriptively and inferentially using the Kolmogorov-Smirnov test, intraclass correlation coefficient (ICC) and MannWhitney test (α=5%). The mean age of the sample was 25.9 years (± 4.7), and 50.9% (n=59) were males. When the coordinates were evaluated, a low correlation was obtained between the images (ICC<0.4). Of the 27 proposed measures, 77.7% (n=21) indicated agreement between the images in the two views (p>0.05). A comparison of ls-g, saec, pa-ec, slb-ec, sba-sa and slb-sa showed disagreement between the images. Therefore, there is agreement between the facial measures in the frontal and lateral images, except for ls-g and for the distances between the ear landmarks.


Introduction
The individual characteristics of the human face facilitate identification of individuals (Tistarelli;Bicego;Grosso, 2009;Moreton;Morley, 2011). Certain techniques, such as anthropometry by means of bone landmarks and photogrammetry and cephalometry by means of photographic and radiographic images (Allanson, 1997), can be applied to craniofacial analysis.
Photoanthropometry is a relatively recent area of study that arose from the need to develop methods for anthropometric analysis of images. It is defined as a method of facial comparison that uses a metric approach via the distances between anthropometric points marked in facial photographs (Moreton;Morley, 2011). Thus, the method quantifies the characteristics and proportions of an individual based on the landmarks, angles and dimensions of their face in images (İşcan;Helmer, 1993;Porter;Doran, 2000).
This method is an important tool for estimating age, assessing sex and comparing facial patterns (Farrera;García-Velasco;Villanueva, 2016;Gonzalez;Michel-Crosato, 2018). Studies show that indices obtained from facial measurements in photographs show a strong correlation with the real measurements of the individual (Cattaneo et al., 2012).
In forensics, it is common for only images, in photographs or videos, to be available as a source for identifying subjects. Therefore, the photoanthropometric method needs to be studied so that it can be used to prevent and combat crimes that spread rapidly across the world, such as child pornography (Berkowitz, 2009).

Methodology
This is a retrospective documentary study, with primary data obtained from an unwritten source (Lakatos;Marconi, 2021), in which facial photographs, in frontal and lateral views, of the same individual were analyzed. It was submitted to and approved by the Research Ethics Committee under CAAE: 67264117.9.0000.5188.
All the images were exported from a civil database and randomly distributed in Two-dimensional Forensic Facial Analysis System software for frontal and lateral images, SAFF-2D ® and SAFF-2D Profile ® , developed by the National Forensics Institute of the Brazilian Federal Police.
Of the 188 pairs of images, 72 pairs were excluded due to the inability to view all the photoanthropometric landmarks proposed in this study. Therefore, the sample comprised 116 pairs of facial images of Brazilian individuals between 20 and 40 years old with equal proportions of sexes.
The photographs of the two views were then scaled by a single examiner, according to size, using CorelDRAW X6 ® software (Corel Corporation, Canada). For this purpose, the ectocanthion (the most lateral point of the eye) and superaurale (the uppermost point of the outer ear) were used as reference points. The rotation of the profile image was adjusted using the subaurale (lowest point of the earlobe) or subnasale (lowest point of the nose) as a reference, as shown in figure 1, depending on which point was more visible in the images of the two views. These reference points were chosen because they are easily visualized in both front and profile images.

Figure 1
Representation from front and side photographs in the process of adjusting size and rotation using the ectocanthion and subnasale as a reference points, using the CorelDRAW X6 ® software.

Source: Authors.
Then, each image was individually analyzed using SAFF-2D ® and SAFF-2D Profile ® , and two examiners marked 16 facial anatomical landmarks, as shown in Figures 2 and 3, visible in the photographs of the two views by means of the photoanthropometric adaptation proposed by Flores and Machado (2017), which is described in Table 1. Table 1 Photoanthropometric definitions proposed by Flores and Machado (2017) for the 16 facial points used in this study.

Photoanthropometric landmark Definition
Glabella (  Research, Society and Development, v. 10, n. 5, e27610514795, 2021 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v10i5.14795 The intra-and interrater agreement was evaluated to assess the reliability of the markings. Initially, 10 images were marked by each examiner. After 30 days, the marking was repeated for the same photographs, and the correlation between the landmarks was analyzed using the intraclass correlation coefficient (ICC) (Shrout;Fleiss, 1979), which was also used to analyze the agreement between the coordinates of the landmarks in the front and profile images.
Then, the landmarks were identified in the sample, and the pixel coordinates of the respective x and y-axes were obtained. In addition, 27 vertical distances between the landmarks were proposed, as follows: ls-sn, n-g, sn-g, ls-g, sto-g, li-g, gn-g, sn-n, ls-n, sto-n, li-n, gn-n, sto-sn, li-sn, gn-sn, sto-ls, li-ls, gn-ls, li-sto, gn-sto, gn-li, sa-ec, sba-ec, pa-ec, slb-ec, sba-sa and slb-sa. For standardization of the vertical measurements, the values corresponding to the y-coordinates were corrected.
Initially, for each photoanthropometric landmark, the mean value of this coordinate was taken, which was defined as the correction factor. Subsequently, the correction factor was subtracted from the real coordinate value, and the corrected value was obtained. For the calculation of the vertical distances, the corrected values of the landmarks were subtracted for each proposed measure.
The data were analyzed in a descriptive and inferential manner using Microsoft Excel ® 2013 and IBM SPSS ® v.21 software. The images were initially compared using descriptive data, with the mean and standard deviation. The parameter normality was assessed using the Komolgorov-Smirnov (KS) test. The ICC was estimated in two distinct ways, single and average, using the absolute agreement as a measure. Cronbach's alpha (Cronbach, 1951) was also calculated to measure the level of agreement between the two markings in the frontal and lateral views because, compared to the ICC average, this term Research, Society and Development, v. 10, n. 5, e27610514795, 2021 (CC BY 4.0) | ISSN 2525-3409 | DOI: http://dx.doi.org/10.33448/rsd-v10i5.14795 8 assesses the existence of bias (Vargo, 2003). The Mann-Whitney test was applied to compare the vertical distances of the groups of frontal and profile images.

Results
The intrarater evaluation, which together with the interrater evaluation preceded the marking of the photographs of the sample, indicated low dispersion (≤1 mm), with agreement in 78.1% (n=25) of the landmarks and ICC˃0.8.
For all 16 photoanthropometric landmarks, low agreement was observed between the x-and y-coordinates obtained in the front and profile images, as all exhibited ICC <0.4, with ICC<0.1 in 25% (n=4), 0.1<ICC<0.2 in 68.8% (n=11), and ICC>0.2 in 6.2% (n=1). A comparison of the ICC average with Cronbach's alpha showed that because the values for both indices were equal, as shown in Table 2, indicating that there were no systematic errors. In the table above, the ICC values call attention, which were low for all the points studied. This indicates that, in pairs of photographs, it is not recommended to analyze exclusively the values referring to the coordinates of the points, as they have low concordance with each other.
Of the 27 proposed measures, 77.7% (n=21) indicated agreement between the images of the two views (p>0.05).
Only for distances ls-g, sa-ec, pa-ec, slb-ec, sba-sa and slb-sa was there disagreement between the images, as shown in Table 3. of the proposed measures, 21 are corresponding between the images of the same individual from the front or in profile.

Discussion
The facial analysis of photographs for the identification of criminals originated in the nineteenth century, from a method called Bertillonage based on body and facial measurements and morphological evaluations, and was incorporated by the American and British police in the early twentieth century Evans, 2009).
Since the advent of monitoring equipment for safety purposes, such as closed-circuit television (CCTV), circa 1985, in the United Kingdom Evans, 2009) and the rapid growth in the production of images, facilitated by the evolution of mobile phones, develop and/or improving identification methods from photographs or videos has been a goal in forensics.
Despite its relevance, there is still a shortage of publications in this area, including those regarding reliable methodologies of analysis, which leads to questions about the validity of the method (Alves et al., 2021) and, therefore, the forensic work performed based on these.
In this sense, Baldasso et al. (2016) stated that from reference points in distorted images, it is possible to obtain the ideal positions of these points by correcting the planar projection using frontal photographs as references for the images at different angles, similar to the method used in this study.
The facial comparison -or mapping -involves the description of the characteristics of a population, as well as the identification of individuals, and is applicable as evidence in various societies around the world (Akhter et al., 2013;Davis; Valeitine; Davis, 2010;Roelofse;Steyn;Becker, 2008;Ogawa et al., 2015). Studies seeking to evaluate facial growth or estimate age from landmarks in frontal photographs have achieved positive results with this method (Cattaneo et al., 2012;Borges et al., 2018;Machado et al., 2017), although they emphasize that each population requires a specific approach (Cattaneo et al., 2012;Machado et al., 2019).
The scaling technique aims to align the photographs and ensure that they match, enabling the images to be objectively analyzed and compared using quantitative procedures (Baldasso et al., 2016). For the alignment between pairs, three of the sixteen points were used; however, when comparing only the x-and y-coordinates of each point in the images in the two views, a low correlation was observed.
A study on photographs of male subjects of three European nationalities proposed 24 absolute measurements and 24 head and face indices with the goal of establishing a biological profile for identification purposes. The results showed that except for the labial width and intercanthal-mouth index, there were significant differences between the features of the three populations studied (Ritz-timme et al., 2011). These results are in line with the results described here, which showed agreement between most of the vertical distances proposed for comparing the faces of individuals photographed in two different views.
This is an important finding for forensics because it is common to have to use nonfrontal-view images as a source of information. Therefore, it is recommended that greater attention be paid to facial measurements because the vertical distances of a face in profile can be equivalent to those of the frontal view. In addition, the comparison of landmarks only based on the coordinates did not yield good results.
It should be noted that the data presented here represent a small sample of only young adults. Studies that comprise a larger population and other age groups should be developed, in particular to ensure that this is a useful method of analysis in the investigation of cases of child pornography, where knowledge of the age of the individual represented in the image is essential for proving the materiality of the facts and deciding whether to escalate the punishment of the offender (Ferguson; Wilkinson, 2017).

Conclusion
In conclusion, photoanthropometry is applicable as a method of comparison between photographs of the same individual in frontal and lateral views because there is agreement between the observed vertical facial measurements. It should be emphasized that distances involving the ear landmarks are not yet reliable for comparison, and caution must be taken when adjusting the photographs to align the landmarks and standardize the measurements between the images in the two views.
It is suggested the continuity of research in more diverse samples, with images of individuals in profile and in different angles in order to confirm the result obtained from the proposed indexes adapting them to the forensic purposes.