Assessment of presentation methods for ReFace computerized facial approximations

Cranial and facial inter-landmark distances and tissue depth dataset from computed tomography scans of 388 living persons

Computed tomography (CT) scans of 388 living adults of both sexes were collected from four self-identified ancestry groups from the United States (African, Asian, European, and Hispanic). Scans were acquired from multiple institutions and under a variety of scanning protocols. Scans were used to produce 3D bone and soft tissue models, from which were derived cranial and facial inter-landmark distances (ILDs) and soft tissue depth measurements. Similar measurements were made on 3D facial approximations produced by ReFace software. 3D models and all measurements were obtained using MimicsR software. These measurements are useful for facial approximations of unidentified decedents and for investigations into human variation between and among ancestry groups and sexes.

3D forensic facial approximation: Implementation protocol in a forensic activity

The primary objective of this paper is to report on the successful implementation of forensic facial approximation in a real case in the forensic context. A three-dimensional (3D) facial approximation protocol of the skull was performed with free software, applying techniques in a virtual environment that have already been consolidated in the literature. The skull was scanned with the photogrammetry technique, the digital replica was imported in the Blender software (Blender Foundation, Amsterdam) and individualized model sketches of the face were traced with the MakeHuman software (MakeHuman Org) according to the anthropological profile of the victim. The face created was imported in Blender, where it was adapted, modeled, and sculpted on the 3D skull and its soft tissue markers, using an American open-source application of the technique in the digital environment. The face created in a virtual environment was recognized and legal identification procedures were started, resulting in the more agile delivery of the disappeared body to its next of kin. It is therefore concluded that facial approximation may not be a primary method of human identification, but it can be satisfactorily applied in the forensic field as an individual recognition resource. It has great value in narrowing the search, reducing the number of alleged victims, and leading to identification tests, therefore significantly reducing the number of genetic DNA (deoxyribonucleic acid) tests-which are considered costly for the State or Federation-and consequently reducing the waiting time before delivery of the body to its family.

Recognizability of computer-generated facial approximations in an automated facial recognition context for potential use in unidentified persons data repositories: Optimally and operationally modeled conditions

Currently in the United States, the remains of thousands of unidentified human decedents are housed in medical, law enforcement, and forensic facilities throughout the country. A number of digital data repositories have been established to curate and disseminate the details of these unidentified decedent cases; some repositories also maintain records of missing persons. Although a cross-reference for textual data similarity occurs between the missing persons and unidentified decedent records in some repositories, no repository is currently known to employ an image analysis technology for cross-referencing image data. Results suggest that the computer-generated facial approximations used in this research were consistently included in prioritized candidate lists when used in an automated facial recognition context. Two concurrent studies exploring the specific use-case discussed here were executed. The first employed an optimally-conditioned facial image gallery (g=6159) (i.e., a gallery comprised of highly consistent facial images), a research design intended to establish the ceiling performance of the combined use of the two software programs employed. The second employed a gallery (g=1816) compiled from a real-world dataset of missing persons' facial images, a research design intended to inform potential operational performance when using the highly varied facial images typically comprising public databases. Multiple types of facial approximations (reconstructions) with varying degrees of weight adjustments, age adjustments, or the presence (or absence) of visible eyes, and combinations of these variables, were evaluated. Overall, in the larger, optimally modeled study, 53% of the facial approximations for the t=159 test subjects examined were matched to his or her corresponding life photo within the top 50 images of a candidate list generated from a blind (unrestricted) search of the highly consistent gallery (g=6159). In the operationally modeled study, 31% of the test subjects' (t=16) facial approximations were matched to their corresponding life photos within the top 50 images of a candidate list generated from a blind search of the gallery populated with images from an operational dataset (g=1816). As anticipated, candidate list inclusion rates improved with the use of demographic filters. No significantly different inclusion rates were observed between the sex or age cohorts examined. Significant differences were, however, observed across population cohorts. Entities curating missing and unidentified decedent records may benefit from a paired implementation of facial recognition technology and computer-generated approximations as part of a comprehensive investigative strategy for the specific envisioned use-case discussed in this research.

Assessment of accuracy and recognition of three-dimensional computerized forensic craniofacial reconstruction

Facial reconstruction is a technique that aims to reproduce the individual facial characteristics based on interpretation of the skull, with the objective of recognition leading to identification. The aim of this paper was to evaluate the accuracy and recognition level of three-dimensional (3D) computerized forensic craniofacial reconstruction (CCFR) performed in a blind test on open-source software using computed tomography (CT) data from live subjects. Four CCFRs were produced by one of the researchers, who was provided with information concerning the age, sex, and ethnic group of each subject. The CCFRs were produced using Blender® with 3D models obtained from the CT data and templates from the MakeHuman® program. The evaluation of accuracy was carried out in CloudCompare, by geometric comparison of the CCFR to the subject 3D face model (obtained from the CT data). A recognition level was performed using the Picasa® recognition tool with a frontal standardized photography, images of the subject CT face model and the CCFR. Soft-tissue depth and nose, ears and mouth were based on published data, observing Brazilian facial parameters. The results were presented from all the points that form the CCFR model, with an average for each comparison between 63% and 74% with a distance -2.5 ≤ x ≤ 2.5 mm from the skin surface. The average distances were 1.66 to 0.33 mm and greater distances were observed around the eyes, cheeks, mental and zygomatic regions. Two of the four CCFRs were correctly matched by the Picasa® tool. Free software programs are capable of producing 3D CCFRs with plausible levels of accuracy and recognition and therefore indicate their value for use in forensic applications.

Automated facial recognition of manually generated clay facial approximations: Potential application in unidentified persons data repositories

This research examined how accurately 2D images (i.e., photographs) of 3D clay facial approximations were matched to corresponding photographs of the approximated individuals using an objective automated facial recognition system. Irrespective of search filter (i.e., blind, sex, or ancestry) or rank class (R₁, R₁₀, R₂₅, and R₅₀) employed, few operationally informative results were observed. In only a single instance of 48 potential match opportunities was a clay approximation matched to a corresponding life photograph within the top 50 images (R₅₀) of a candidate list, even with relatively small gallery sizes created from the application of search filters (e.g., sex or ancestry search restrictions). Increasing the candidate lists to include the top 100 images (R₁₀₀) resulted in only two additional instances of correct match. Although other untested variables (e.g., approximation method, 2D photographic process, and practitioner skill level) may have impacted the observed results, this study suggests that 2D images of manually generated clay approximations are not readily matched to life photos by automated facial recognition systems. Further investigation is necessary in order to identify the underlying cause(s), if any, of the poor recognition results observed in this study (e.g., potential inferior facial feature detection and extraction). Additional inquiry exploring prospective remedial measures (e.g., stronger feature differentiation) is also warranted, particularly given the prominent use of clay approximations in unidentified persons casework.

Recognition of computerized facial approximations by familiar assessors

Studies testing the effectiveness of facial approximations typically involve groups of participants who are unfamiliar with the approximated individual(s). This limitation requires the use of photograph arrays including a picture of the subject for comparison to the facial approximation. While this practice is often necessary due to the difficulty in obtaining a group of assessors who are familiar with the approximated subject, it may not accurately simulate the thought process of the target audience (friends and family members) in comparing a mental image of the approximated subject to the facial approximation. As part of a larger process to evaluate the effectiveness and best implementation of the ReFace facial approximation software program, the rare opportunity arose to conduct a recognition study using assessors who were personally acquainted with the subjects of the approximations. ReFace facial approximations were generated based on preexisting medical scans, and co-workers of the scan donors were tested on whether they could accurately pick out the approximation of their colleague from arrays of facial approximations. Results from the study demonstrated an overall poor recognition performance (i.e., where a single choice within a pool is not enforced) for individuals who were familiar with the approximated subjects. Out of 220 recognition tests only 10.5% resulted in the assessor selecting the correct approximation (or correctly choosing not to make a selection when the array consisted only of foils), an outcome that was not significantly different from the 9% random chance rate. When allowed to select multiple approximations the assessors felt resembled the target individual, the overall sensitivity for ReFace approximations was 16.0% and the overall specificity was 81.8%. These results differ markedly from the results of a previous study using assessors who were unfamiliar with the approximated subjects. Some possible explanations for this disparity in performance were examined, and it was ultimately concluded that ReFace facial approximations may have limited effectiveness if used in the traditional way. However, some promising alternative uses are explored that may expand the utility of facial approximations for aiding in the identification of unknown human remains.

Forensic Facial Reconstruction: Relationship Between the Alar Cartilage and Piriform Aperture

During forensic facial reconstruction, facial features may be predicted based on the parameters of the skull. This study evaluated the relationships between alar cartilage and piriform aperture and nose morphology and facial typology. Ninety-six cone beam computed tomography images of Brazilian subjects (49 males and 47 females) were used in this study. OsiriX software was used to perform the following measurements: nasal width, distance between alar base insertion points, lower width of the piriform aperture, and upper width of the piriform aperture. Nasal width was associated with the lower width of the piriform aperture, sex, skeletal vertical pattern of the face, and age. The current study contributes to the improvement of forensic facial guides by identifying the relationships between the alar cartilages and characteristics of the biological profile of members of a population that has been little studied thus far.

Comparison Among Manual Facial Approximations Conducted by Two Methodological Approaches of Face Prediction

This study verified the difference between two methods of forensic facial approximation (FFA) regarding recognition and resemblance rates. Three-dimensional models of skulls were obtained from computerized tomography (CT) scans of two subjects (targets). Two manual FFAs were performed for each target, by applying two different guidelines for the facial structures (what we called "American method" (AM) and "Combined method" (CM)). Unfamiliar assessors evaluated the sculptures by recognition and resemblance tests. The AM was that which allowed more correct responses of recognition and higher resemblance's scores for the male target (p < 0.001). Regarding guidelines for modeling characteristics of the face, the ones that are practical and easily performed for sculptures, such as the length of the anterior nasal spine multiplied by 3 for nose prediction, may offer better results in terms of resemblance.

Forensic facial approximation assessment: can application of different average facial tissue depth data facilitate recognition and establish acceptable level of resemblance?

Facial soft tissue thicknesses (FSTT) are important guidelines for modeling faces from skull. Amid so many FSTT data, Forensic artists have to make a subjective choice of a dataset that best meets their needs. This study investigated the performance of four FSTT datasets in the recognition and resemblance of Brazilian living individuals and the performance of assessors in recognizing people, according to sex and knowledge on Human Anatomy and Forensic Dentistry. Sixteen manual facial approximations (FAs) were constructed using three-dimensional (3D) prototypes of skulls (targets). The American method was chosen for the construction of the faces. One hundred and twenty participants evaluated all FAs by means of recognition and resemblance tests. This study showed higher proportions of recognition by FAs conducted with FSTT data from cadavers compared with those conducted with medical imaging data. Targets were also considered more similar to FAs conducted with FSTT data from cadavers. Nose and face shape, respectively, were considered the most similar regions to targets. The sex of assessors (male and female) and the knowledge on Human Anatomy and Forensic Dentistry did not play a determinant role to reach greater recognition rates. It was possible to conclude that FSTT data obtained from imaging may not facilitate recognition and establish acceptable level of resemblance. Grouping FSTT data by regions of the face, as proposed in this paper, may contribute to more accurate FAs.

Craniofacial Superimposition: Historical Review and Current Issues

Methodology of craniofacial superimposition has evolved dramatically since its inception in the early 20th century. Contemporary approaches involve advanced procedures of digitization, use of video cameras, warping algorithms, skull positioning devices and data derived from computed tomography, radiography, and ultrasound technology. Applications have proven most useful in exclusion and casework frequency has declined in some regions in recent years as molecular approaches to identification have become more widely available. Additional research is needed to clarify the probabilities involved and to facilitate comparisons when the technique is employed.