Security systems rely on detailed images of hands to ensure permissions and access to private information and spaces. So while hands can definitely be used for identification purposes, evidential images of hands pre-, during and post-crime are never of the quality used for infrared imaging on security systems.
Evidence is often captured in uncontrolled situations, making it difficult to analyze. Still, researchers from Lancaster University (UK) argue there is a strong need to investigate the potential for identification from digital images of the hand, especially in cases of sexual assault where hand information may be the only biometric available for analysis.
“Hand images, one of the primary biometric traits, provide distinctive features for person identification. They also have less variability when compared to other biometric modalities, such as faces,” write the researchers.
In their paper published in arXiv, the team—which includes renowned forensic anthropologist Sue Black—proposes a new machine learning framework called the Global and Part-Aware Network (GPA-Net). What makes this method unique from previous work on the topic is the separation of right and left hands in the datasets used to train and analyze.
Previous research has treated both the right and left dorsal (back of hand) or palmar (front of hand) of a subject as the same. But, according to Nathanael Baisa et. al, it is not possible to compare right and left hands since they show disparity in their vein patterns. Instead, for their GPA-Net, Baisa and his colleagues separated the dataset into right dorsal, left dorsal, right palmar and left palmar sub-datasets.
The research team tested the GPA-Net framework across two datasets: the 2016 11k Hands set, a collaboration between researchers from Canada and Egypt, and the Hong Kong Polytechnic University Hand Dorsal (HD) set. The 11k Hands dataset has 190 subjects, while the HD dataset comprises 502 subjects.
In experiments, GPA-Net’s results surpassed two competing machine learning architectures—ResNet50 and VGG19-bn. According to the published paper, GPA-Net outperformed ResNet50 by 25% accuracy and 38% mean average precision (mAP), while it surpassed VGG-bn by 40% accuracy and 38% mAP.
“The experimental results on two public hand datasets demonstrate the superiority of the proposed method over the competing methods, giving it strong potential for robust identification of the perpetrators of serious crime,” concluded the authors.
Photo: From top to bottom row are right dorsal of 11k, left dorsal of 11k, right palmar of 11k, left palmar of 11k and HD datasets. The green and red bounding boxes denote the correct and the wrong matches, respectively. Credit: Nathanael L. Baisa, et. al, arXiv:2101.05260