Computer assisted photo-anthropometric analyses of full-face and profile facial images

https://doi.org/10.1016/j.forsciint.2010.04.012Get rights and content

Abstract

Expert witnesses using facial comparison techniques are regularly required to disambiguate cases of disputed identification in CCTV images and other photographic evidence in court. This paper describes a novel software-assisted photo-anthropometric facial landmark identification system, DigitalFace tested against a database of 70 full-face and profile images of young males meeting a similar description. The system produces 37 linear and 25 angular measurements across the two viewpoints. A series of 64 analyses were conducted to examine whether separate novel probe facial images of target individuals whose face dimensions were already stored within the database would be correctly identified as the same person. Identification verification was found to be unreliable unless multiple distance and angular measurements from both profile and full-face images were included in an analysis.

Introduction

In recent years the interrelated topics of identity and surveillance have regularly been the features of the political and news agenda. Widespread instances of identity fraud have led to the implementation of preventative measures such as biometric passports and identity cards. Additional crime reduction initiatives have also resulted in the increasing presence of Closed Circuit Television (CCTV) systems. Both types of image are often used within the criminal justice system for identification purposes, with cases resting on whether a suspect is the individual depicted. Offenders often confess if presented with this type of evidence. However, when cases depend on disputed identification in photographs, there are various approaches to resolving the issue, and in the UK specific guidelines have been issued by the Attorney General [1].

Witnesses who know a defendant personally can testify as to identity from photographic evidence and research has regularly found that recognition of highly familiar faces in CCTV images is reliable, even if quality is poor (e.g., Refs. [2], [3]). Juries can also be encouraged to conclude that images depict the defendant and in the UK, if images are ‘sufficiently clear’, no other form of identity evidence is required. Jurors and most police officers will be unfamiliar with a suspect and even with high-quality images and with no memory demands, the simultaneous identification matching of unfamiliar people is susceptible to error, resulting in high levels of confident false positive and negative decisions [2], [4], [5]. Furthermore, the presence of the target in person does not increase the safety of this type of matching judgment [6], [7].

Experts in face structure may also provide opinion evidence of identity from examination of evidential images. Commonly known as facial comparison or facial mapping, this type of evidence is admissible worldwide (e.g., Australia: Ref. [8]; India: Ref. [9]; Italy: Ref. [10]; South Africa: Ref. [11]; USA: Refs. [12], [13]) and in the UK more than 500 court reports of this type are prepared per annum [14]. Practitioners often use a combination of three approaches to facial comparison.

With morphological comparison, facial features are graded by shape and size into discrete categories (e.g., Refs. [10], [12], [15]). It is probably the most effective technique with images of poor quality or from different viewpoints, as prominent individuating features remain visible. However, features may possess elements of more than one category, are on the boundary between two, or are unclassifiable. Furthermore, statistical analyses can only be conducted at a nominal level, meaning that discrimination of faces possessing similar characteristics will be problematic. Furthermore, features that successfully discriminate one population from a specific geographical region may not individuate those from another [16].

With photo-anthropometry (e.g., Refs. [17], [18], [19], [20], [21]), proportional analyses of the distances and angles between anatomical landmarks in images are calculated and compared. Precise values can be acquired, allowing confidence assessment and parametric analyses. However, problems are encountered if images are not aligned or facial expressions differ, and therefore a combination of photo-anthropometry and morphological comparison techniques may be of most use for examining two facial images [11].

Finally, with superimposition (e.g., Refs. [22], [23], [24]), one image is projected over another in order to highlight facial similarities, or more saliently, discrepancies as well as the possession, or the lack of facial symmetry. However, disambiguation can be hindered by viewpoint differences or style of presentation such as a slow fade. Superimposition using three-dimensional (3D) images may be most successful when used in conjunction with anthropometry and morphological comparison. Yoshino and colleagues [23], [24] demonstrated viewpoint-matched superimposition of 3D facial images over traditional camera images, allowing accurate identity judgments. Anthropometrical locations were marked on both images and software automatically extracted a two-dimensional image from the 3D image for alignment with the camera photograph. Using reciprocal point-to-point distances, the system was found to be 100% accurate at matching a database of 100 disguised photographed faces with 3D images of the same individuals. Other researchers (e.g., Ref. [25]) have found comparable rates of success using 3D laser scans.

Regardless of method, only proof of non-identity is possible. A single reliable difference has more weight than a multitude of similarities. Even if results suggest that two different images are of the same person, proof of identity cannot be established as fact. Furthermore, CCTV images are often of poor quality, with feature boundaries indistinct, obscured, in shadow or distorted. Posed identity photographs may not suffer from these problems, although source equipment discrepancies, brightness range, colour capture and reproduction can alter appearance. Camera and lens quality, distance, the nature and direction of lighting, blur, pixel resolution with digital images, and analogue tape quality must also be considered. Nevertheless, the greater the number of similarities between two images, the greater the likelihood of an identity match.

Given these issues, it is perhaps not surprising that facial comparison has been criticized, particularly as two or more witnesses using similar techniques can come to opposite conclusions [14], [26]. In the UK, judges [27] have called for the establishment of large-scale facial measurement databases to calculate the likelihood that two different individuals possess similar face structures. Without this safeguard the court ruled that opinions could be regarded as subjective. However, practitioners often apply their techniques using hand-held equipment such as calipers and constructing such a database might be impractical.

Extensive literature has been published describing automatic face identification algorithms (e.g., Refs. [28], [29], [30]), some being more accurate than humans under optimal conditions at correctly identity matching or discriminating two different faces when tested against large databases. However, when viewpoint is incongruent, or images are filmed under different environmental conditions, accuracy of computer facial recognition systems is worse than human performance [31], [32]. Future innovations may prove successful in eliminating or identifying potential suspects. However, human observers will probably continue to personally examine evidence before testifying in court.

Some photo-anthropometrical measures have been published (e.g., Refs. [17], [18], [19], [20], [21]), as have reports detailing its successful application in real criminal trials (e.g., Refs. [8], [12]). Nevertheless, this has mostly involved a limited set of measurements against nonexistent, or small facial databases, containing heterogeneous groups of people unlikely to be the subject of mistaken identifications. Kleinberg et al. [20] analyzed distance and angular measurements derived from a set of four landmarks on a database of 120 male police recruits, concluding that the use of this limited measurement set was less reliable for face discrimination than the application of morphological comparison techniques.

Mardia et al. [18] analyzed a database of 358 young adult white male faces, captured in full-face and profile views taken in a controlled environment. Twenty distance and angular measurements between landmarks were collected. High correlations between all measurements highlighted limitations in face discrimination even with these extremely high-quality viewpoint-standardized images. Full-face and profile viewpoint analyses were conducted separately and if combined, as might be possible if multiple images were acquired in a real legal case, a more robust method may have emerged. Nevertheless, several successful prosecutions have been secured in Australia using a combination of techniques, including fewer than ten anthropometrical measurements [8].

For the current paper, a custom software-assisted facial landmark photo-anthropometric identification system, DigitalFace was designed, and was evaluated to examine whether its application could be useful to the courts. With images displayed on a computer monitor, operators identify and save up to 38 facial landmark sites in full-face (anterior) view; and 14 landmarks in profile view, producing a database of 37 linear and 25 angular measurements, more than the number used in previous published research. Repeated attempts by the same or different operators for reliability analyses are possible and DigitalFace can be used with any images taken from the same viewpoint, although it was specifically designed for full-face and profile images, such as police ‘mug-shots’. Minimum acceptance criteria for identity determination were generated, tested against a database of 70 male faces meeting a similar description.

Reliability of the system to identify two images (probe and target) as the same person and not to match a probe with a distracter in the database was examined. Match or mismatch judgements between items in the database were based on a converted distance in Euclidean space. The Euclidean space is an implementation of the concept of a face-space, in which the parameters derived by DigitalFace form the dimensions of the space [33], [34]. Perceptual measures of facial similarity were also acquired of all pairs of database faces. Some had been prone to incorrect identification as the same person by human participants in previous matching experiments [6]. These experiments replicated scenarios that might occur when CCTV evidence has been acquired. Thus, circumstances were similar to cases in which expert witnesses might be requested to assist in determining identity.

There were two criteria for a positive identification decision. Firstly, the probe and target measurements (proportional distances and/or angles) of two images of the same person should be closer in Euclidean space than the distance to any distracter. The second, more rigorous criterion was that the distance between two images of the same person should additionally be less than that between all other pairs of images of two different faces. The latter meets court recommendations for testing against a relevant database to provide an estimate of the likelihood of other members of the population possessing similar attributes. The probe images were of eight young males taken three weeks after their target database photographs had been acquired. Hairstyle had slightly changed and there were minor viewpoint variations as might occur naturally even with posed passport photos. Features may be obscured or unclear in CCTV available from a crime scene and the specific choice of measurements in any facial comparison analysis will depend on availability. Therefore, a series of 64 separate analyses were conducted using alternative measurements, including angular only to assess viewpoint consistency.

Section snippets

Materials

The full-face and profile probe and target photographs were of eight volunteers recruited to a previous identity matching study [6]. All were male white European (aged 18–22), of slim or medium/muscular build (72–95 kg, 1.70–1.92 m), clean-shaven, with brown or black hair, neither receding, nor below collar length. None had distinguishing facial marks or wore jewelry. The probe images were taken approximately three weeks after the target images. Photographs of 62 distracters meeting a similar

General interpretation of photo-anthropometrical analyses

A series of 64 independent analyses were conducted. Each of the eight probe faces were separately included with the database faces (which included the target) to generate a proximity matrix, using eight different sets of measurements as might be necessary with evidence acquired in the course of a criminal investigation (Table 2, Table 3, Table 4, Table 5, Table 6, Table 7, Table 8, Table 9). Alternative measures might be preferred in different circumstances. Reported in each table is the

Discussion

The findings of the photo-anthropometrical analyses using the DigitalFace landmark identification system illustrate that caution is required if deciding whether two different photographs depict the same person. Even when all distance and angular measurements were included in full-face view, and separately with all measurements in profile view, there were a number of failures to match on the robust secondary criteria that the Euclidean distance between a target and a probe should be less than

Acknowledgements

This work was funded by a Postgraduate Studentship Grant PTA-030-2002-00990 from the Economic and Social Research Council, UK. Part of this research was presented at the British Psychological Society Annual Conference, University of York in March 2007. We thank two anonymous reviewers for their helpful comments on an earlier version of this work.

References (36)

  • R.J. Allen

    Exact solutions to Bayesian and maximum likelihood problems in facial identification when population and error distributions are known

    Forensic Sci. Int.

    (2008)
  • Attorney General's Reference No. 2 of 2002, EWCA Crim 2373, UK,...
  • V. Bruce et al.

    Matching identities of familiar and unfamiliar faces caught on CCTV images

    J. Exp. Psychol.: Appl.

    (2001)
  • A.M. Burton et al.

    Face recognition in poor-quality video: evidence from security surveillance

    Psychol. Sci.

    (1999)
  • V. Bruce et al.

    Verification of face identities from images captured on video

    J. Exp. Psychol.: Appl.

    (1999)
  • Z. Henderson et al.

    Matching the faces of robbers captured on video

    Appl. Cogn. Psychol.

    (2001)
  • J.P. Davis et al.

    CCTV on trial: matching video images with the defendant in the dock

    Appl. Cogn. Psychol.

    (2009)
  • R. Kemp et al.

    When seeing should not be believing: photographs, credit cards and fraud

    Appl. Cogn. Psychol.

    (1997)
  • Cited by (44)

    • Physiotherapeutic methods of treatment of mandibular distal occlusion and the progress of therapy: A case report

      2022, Journal of Bodywork and Movement Therapies
      Citation Excerpt :

      The examination yields precise results and helps assess the progress of therapy, but it is invasive and cannot be repeated often, especially among children in developmental age (Pakbaznejad Esmaeili et al., 2016; Goodwin et al., 2017). The paper presents a non-invasive photoanthropometric method (Stengel-Rutkowski et al., 1984; Hung et al., 2004; Davis et al., 2010; Flores et al., 2019; Graja et al., 2019; Menéndez López-Mateos et al., 2019). The necessary instruments are readily available and measurements are easy to make (Farhud et al., 2001).

    • Human identification by the ear: Reproducibility and applicability in a Brazilian sample

      2021, Forensic Imaging
      Citation Excerpt :

      In this context, images of good quality may comprise the only available and suitable material to be used for human identification [1,2]. As they have been also described as a useful tool for the exclusion of suspects in crime scene investigation [2], photographs of the ear can also be used in the post mortem identification as it is a body region that has individualizing characteristics that can remain preserved for some time after death [3–5], although its application depends on the quality of the material available for analysis [6–10]. Therefore, the use of ear images has been the subject of the research that employs automated analysis [11–14].

    • Computer-Assisted Anthropology

      2019, Internet of Things in Biomedical Engineering
    View all citing articles on Scopus
    View full text