Biometric recognition of the hand in unconstrained images
Publication Date
July 27, 2026
Creator
Hornshaw, Gabrielle
Abstract
The aim of this thesis is to progress the use of hand recognition in forensic contexts to aid in intelligence gathering and identification in perpetrator-recorded images depicting criminal activity (such as sexual abuse). Most developments in hand recognition to date have been evaluated on highly constrained hand image datasets with little variation in factors such as hand pose and illumination, which poorly represents the images that would be seen in a forensic context.
In this thesis, a new hand image dataset, named VarHands, with 6436 images of 40 hands from 20 subjects is collected. In it, each hand is featured under four illuminations with large variety in hand pose. A subset of hand recognition methods identified from the literature are evaluated against it to determine which current approaches are best suited to the unconstrained domain. Numerous pre-processing and model training techniques are proposed to improve cross-domain performance when moving from highly constrained training datasets to the unconstrained evaluation data: region-of-interest extraction and matching outperformed whole-hand matching by 3.8% rank-1 accuracy on 11k Hands, and luminance-based illuminationnormalisation achieved a 39.06% rank-1 accuracy improvement on the new collected dataset. Further to this, uncertainty quantification is explored as a way to handle frequent mis-detections, blurring, and occlusion found in unconstrained images, with the proposal of a new uncertainty-aware method of multi-biometric fusion which achieves +0.6% rank-1 accuracy. The results show strong potential of application but highlight the need for more data from which to learn representative distributions of hand features.
The findings from this thesis provide avenues of research to develop hand recognition algorithms that are invariant to hand pose and illumination and thus more suitable for forensic application or other unconstrained domains. To conclude, a review of the regulatory landscape surrounding the application of biometric technologies in law enforcement is performed to identify the requirements for adoption of hand recognition, which further supports the development of larger and more representative hand image datasets.
Item Type
ethesis
Thesis Type
PhD
Supervisors
Keywords
Subjects (LC)
Associated Schools / Departments
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
Gabrielle-Hornshaw-14314359-Thesis.pdf
Size
56.29 MB
Format
Adobe PDF
Checksum (MD5)
5c245b5a463c516dae212af0db0f002c