Introduction

In Denmark it is a criminal offense to have sexual relations with a person under the age of 15 years. If the relation involves the recording or distribution of pornographic material, the age of consent is set at 18 years [1]. In 2008 the Danish National Institute of Public Health conducted a national survey of the health and well-being of adolescents, with a special focus on sexual abuse and violence in childhood and early adolescence [2]. The survey stated that 13.8 % of the respondents reported unwanted sexual experiences with peers and/or adults while they were still under the legal age of consent. A little over a third of these described the unwanted sexual experiences as abuse.

Photographic records of the abuse are sometimes kept by the perpetrator, and sometimes shared with others, e.g., via the Internet. However, the perpetrator usually consciously avoids recording their face making direct identification impossible. In such situations an investigator has to use other details gathered from the video/photographs to establish the identity of the perpetrator. Examples of such details include objects and locales that are connected to a suspect, the identification of a victim who can then be questioned if identified, or the formulation of a forensically based expert opinion on the significance of certain anatomical structures visible in both the video/photographs and on a suspect [3–5]. In the latter case, the hand of the perpetrator may be visible as he manipulates the victim (while holding the video camera with the other hand). Thus, the back of the hand, including vein patterns, may be visible and used for matching with photos of the back of the hand of one or more suspects.

There are only a few publications where a database of photographs of the back of the hand has been used for anthropological identification [5–7]. We aimed to investigate the success rate in matching low-quality images of the right hand with high quality images from the same person, especially the risk of false positive identification when using a restrictive algorithm for image matching. This experimental setup was chosen to match actual case work.

Methods

51 males, all over the age of 18 years, volunteered to have their right hand video recorded and photographed. All participants were light-skinned and were asked to remove rings, watches, or other jewelry. Individuals with very large scars, amputated digits, and/or other easily identifiable characteristic disfigurations were excluded.

For video recording a Canon Legria HFR37 digital camcorder with a resolution of 1280 × 720 pixels was used, as it represents an average consumer camcorder of a type that has been readily available in recent years. Video recording was carried out using natural lighting. The participant was instructed in a simple exercise where they manipulated a white cylinder, rotating it 360° through six repeated movements. This was to simulate the hand being filmed from different angles, as is often the case in the footage that is under investigation [3]. Still frames with maximum visibility of anatomical structures, scars, and other recognizable features were then selected from the footage.

For high quality photography of the hand a Canon EOS 550D digital single-lens reflex camera with a 17–40 mm photographic objective was used. The photos had a resolution of 18 megapixels, which is comparable to those used by police photo technicians. Photographs were taken under different lighting conditions to ensure a variety of shadows cast by skin creases over the joints of the hand. The hand was placed flat on the surface, but not fully stretched out, during photography. The camera was mounted on a mechanical arm above the hand.

Photo–video comparison and match-determining algorithm

All 51 high quality pictures and 51 video recordings of the volunteers were blinded to the investigator comparing the images. During visual analysis of the images, a distinction between permanent and non-permanent landmarks was made.

Permanent landmarks were defined as:

  • Matching vein pattern, i.e., a very distinct course of a vein, atypical branching patterns, or very large veins.

  • Freckles and nevi, particularly with a diameter over 2 mm and/or any found in very distinct locations, such as the interdigital groove, on top of a finger joint and/or any arranged in a distinct pattern, such as a triangle.

  • Scars, especially larger scars with marked discoloring, as many smaller and paler scars did not show up on the video images.

Non-permanent landmarks were defined as:

  • Body hair, which in theory could be removed or changed between recordings, although in our study any hair growth was naturally the same on both images, since they were obtained at the same time.

  • Nail size and outline, for the same reason listed above.

  • Newer wounds in the process of healing. Although one could hypothesize that wounds, at least larger ones, would lead to the formation of scar tissue, it would be almost impossible to make conclusions about anything other than the location of this scar. Shape, color, and orientation of scarring may vary greatly from person to person, as it is dependent on both genetic factors and the forces acting upon the forming scar during the healing process [8].

By comparing each of the 51 photographs with each of the 51 videos, a total of 2601 (51 × 51) photograph-video image pairs were generated. Of these, 51 pairs are correct matches, i.e., coming from the same person, and 2550 are incorrect. Each photo–video image combination was analyzed and categorized using a set algorithm (Fig. 1).

Fig. 1
figure 1

The algorithm applied by the investigator examining the 2601 possible image combinations in order to classify each image-pair

Full size image

For a given video still frame (x) compared to a given high quality photo (y), the video still was first examined, looking for the presence of any of the three aforementioned types of permanent landmarks. Any landmarks were then sought out in the high quality photo. If a landmark appeared in the still frame but not in the high quality photo, then x and y was categorized as a “non-match,” since any visible permanent landmarks in the low quality still frame should be expected to also be visible in the high quality photo. If this was not the case, the high quality image was then assessed for how many landmarks appeared in the same position on the hand in both pictures. If the number of landmarks that were identified in both pictures was equal to or greater than two, x and y were categorized as a “highly possible match.” If the number was lower than two, x and y were categorized as a “possible match.”

Results

Figures 2, 3, and 4 demonstrate an image pair that was categorized as a “highly possible match.” Figures 5 and 6 demonstrate a false positive “possible match,” i.e., an image pair coming from two different persons but analyzed as possibly being of the same person.

Fig. 2
figure 2

Still frame image taken from one of the video recordings

Full size image
Fig. 3
figure 3

High quality photograph of the same hand as is shown in Fig. 2

Full size image
Fig. 4
figure 4

The same image as shown in Fig. 3 with various distinguishing features circled. Notice in particular the large nevus on the third finger, the small oval-shaped scar at the base of the third finger, and the larger circle-shaped scar on the radial half of the base of the hand

Full size image
Fig. 5
figure 5

Still frame analyzed as a “possible match” with the photograph shown in Fig. 6

Full size image
Fig. 6
figure 6

High quality photograph matched with Fig. 5 because of the single matching permanent landmark (the freckle located along the tendon of the extensor pollicis longus) and no conflicting landmarks. This image was later determined to be of a different person, i.e., a false positive

Full size image

As shown by Table 1, the video still frame from 46 of the 51 individuals had only one “highly possible” matching high quality photo. All of these 46 image pairs were established as being of the same individual, i.e., true positive identifications. There were no “highly possible” matches of images from two different persons.

Table 1 The number of highly possible, possible, and non-matches established to be from either the same person or two different persons
Full size table

In five cases the still frame had no “highly possible match” with any high quality photo as there were no, or only one, permanent landmark identified. In all of these five there was at least one “possible match.”

27 individuals provided a still frame that had multiple “possible matches,” usually two or three. The most uncertain case was one individual, who had five “possible matches.” In total 55 image-pairs were analyzed as “possible matches,” of which 50 were false positives, i.e., the two images came from two different people.

The remaining 2500 possible image pairs were categorized as “non-matches.” There were no false exclusions, i.e., images from the same person analyzed as “non-matches.”

Discussion

If the data is analyzed using only the 46 correct “highly possible” matches being true positives, this corresponds to a negative predictive value of approximately 98 %. As such our method is best suited for ruling out suspects, while a positive match could be used as part of a body of evidence used in the identification of a suspected perpetrator, rather than as conclusive evidence of their identification. This is also what we would expect when applying a restrictive matching algorithm.

We have presented a method of photography using standardized recording conditions for obtaining high quality reference images. In an actual case, the nature of the necessary reference images for comparisons would depend on the evidence material obtained. One example of a case where photographic comparisons were used to identify a suspect revolves around a pornographic recording involving a child, where the only recording of the perpetrator was a brief zoomed in still of the right thumb, which had a distinctive scar. The scar was also present on the thumb of a suspect (already known to the police because of previous offenses). The suspect’s thumb was then photographed at high resolution, along with the thumb of 49 police officers, to demonstrate that the scar in question was only present on the suspect’s finger and not on any others from a random group of Caucasian males [3]. In our setup, the back of the thumb was not included in the reference image database. This demonstrates that a database of reference images would need to be

  • Versatile, containing images of various body parts (particularly arms, hands, and fingers).

  • Taken from different angles (mirrors have been used effectively to achieve this from a single photo [6]).

  • High resolution so that as many anatomical landmarks as possible are visible.

  • Preferably taken under various lighting conditions.

  • Recorded under standardized conditions so all reference images differ only in what is shown and not how it is shown.

This is to ensure that images from the reference database can be meaningfully compared to any body part visible in evidence material. Mounting the camera in a rig or on an arm, using mirrors to show more than one dimension in the same photo, and standardizing hand positioning using pegs, are some suggestions for optimizing recording conditions. The recording equipment that is used should also be accessible at a later date to enable comparable images of possible suspects to be added to the reference database.

The size of the database should be as big as possible, since the point of the comparison is to provide as great an estimate as possible for the hypothesis that a matching image combination is because the images are indeed of the same person and not because of chance. In essence: a more uncharacteristic hand requires a larger database of reference images to produce evidence carrying enough weight to be satisfactory in a courtroom setting. In our study, 24 of the 51 volunteers had sufficiently distinctive features that all other possible image combinations could be ruled out, meaning the evidence would be approximately 50 times more indicative of the person on the high quality picture also being the one on the video. One image had 5 possible matches, meaning the evidence carried about 5 times less weight. As a reference, the British Association of Forensic Science Providers suggests that a probability of evidence, given a proposition (e.g., “the two images are from the same person”), with a value of 100–1000 should be equated to the verbal phrase “moderately strong support” [9]. As such, even for a person having only one matching image, our database should include at least 100 persons to be able to satisfy this requirement.

When presenting results like these in a courtroom setting we recommend developing a likelihood-ratio framework [5, 10, 11]. Using this framework, a case where there is one match (and no other possible matches) carries greater evidential weight than a case where there are several possible matches. This quantifiable outcome can then be integrated into the other available evidence to provide a single fraction comparing the likelihood of the identity of the perpetrator and suspect being the same person, given the presented evidence, to the likelihood that they are two different persons.

Key points

  1. 1.

    Some cases of sexual abuse involve pornographic recording with video and/or regular photography.

  2. 2.

    These images can be used as evidence if a possible perpetrator is apprehended, since it is possible to compare images of that person with the found images.

  3. 3.

    We present a visual-only approach to comparison relying on different anatomical features of the back of the hand for identification.

  4. 4.

    In a sample of 51 males we were able to correctly match all blinded photos to blinded video recordings of the same hand without any false positive identification.

  5. 5.

    We believe this approach is valuable as a supplement to other evidence material, but weak as a standalone method.