Abstract
The analysis of knuckle creases is part of the multifactorial assessment of digital images of the hand used to assist in the identification of perpetrators captured in images depicting child sexual abuse and other offending behaviours. To quantify the impact of finger flexion on the appearance of the dorsal knuckle creases associated with the proximal interphalangeal joint (PIP) joint in digital images, the collection of knuckle crease images, at different points of finger flexion, was facilitated through an app-based Citizen Science project, Knuckle Down ID. A method of knuckle crease classification was adapted to assess the images collected, was evaluated for intra- and inter-reliability, and then used to assess the impact of finger flexion on the frequency of different knuckle crease features observed in manual analysis.The results show that the adapted methodology had good intra-observer repeatability (ICC 0.81) when the entire method was assessed. The assessment of both intra- and inter- observer reliability indicated that the tracing aspect of the methodology was less repeatable than the second half of the methodology; coding the presence of knuckle crease features. It was observed that this had a knock-on impact on the reliability of the whole methodology (ICC 0.50). However, the implementation of a UNet based knuckle crease segmentation algorithm, robust to different hand positions, appeared to show promise in improving the repeatability of the methodology when the outputs of the models were compared to the outputs generated from manual knuckle crease traces produced by experts (ICC 0.71 and 0.61).
The impact of flexion on the recorded frequency of knuckle creases was shown to be significant (p = .00). However, when comparing two hands at different points of flexion, only when comparing fingers held at 0⁰ vs 45⁰ (p =.04) as well as fingers held at 90⁰ vs any other finger position was the differences in the knuckle crease feature frequencies significant (p = .00). The comparison of knuckle crease image pairs at 20⁰ of hyperflexion vs 0⁰ (p = .10) and 20⁰ of hyperflexion vs 45⁰ (p = .09) showed that the difference in knuckle crease feature frequencies was not significant.
This research displays the benefits of integrating a knuckle crease segmentation algorithm to improve the speed and consistency in which the manual analysis of knuckle creases is performed. The results regarding the impact of finger flexion have implications for examiners and improves understanding of the limitations of the 1:1 comparison of knuckle creases in casework evidence in which perpetrator's hands are unconstrained.
| Date of Award | 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Sponsors | Leverhulme Trust |
| Supervisor | Lucina Hackman (Supervisor) & Manuel Trucco (Supervisor) |