Introduction
Security systems worldwide use hand images for biometric identification, ensuring access to sensitive data and restricted spaces. However, forensic evidence involving hands often lacks the controlled quality seen in security systems, making it harder to analyze. In light of this, researchers from Lancaster University, including esteemed forensic anthropologist Sue Black, explored the potential of hand biometrics for robust identification, especially in sexual assault cases where other biometrics might not be available.
The Role of Hand Biometrics in Forensics
Hands offer unique biometric traits with less variability than faces, making them valuable for personal identification. Features like vein patterns provide distinct markers that remain stable over time, adding to their forensic utility.
The study notes that hand information is particularly relevant in scenarios where traditional biometrics, such as facial recognition or fingerprints, are unavailable due to environmental or technical challenges.
Introducing GPA-Net: A Novel Framework
The research team developed Global and Part-Aware Network (GPA-Net), a machine learning model tailored for analyzing hand biometrics. The innovation lies in treating right and left hands as separate datasets to account for their distinct vein patterns, a departure from prior studies that treated both hands as identical.
Dataset and Analysis
- 11k Hands Dataset (2016): A collaborative dataset with data from 190 subjects, sourced from Canada and Egypt.
- Hong Kong Polytechnic University Hand Dorsal (HD) Dataset: Comprising data from 502 subjects.
These datasets were further divided into four subsets:
- Right dorsal
- Left dorsal
- Right palmar
- Left palmar
This categorization ensured more accurate training and analysis, avoiding errors arising from comparing dissimilar hand types.
Performance of GPA-Net
The team benchmarked GPA-Net against two leading machine learning frameworks, ResNet50 and VGG19-bn, demonstrating its superiority in identifying individuals based on hand images:
- Accuracy:
- GPA-Net surpassed ResNet50 by 25%.
- GPA-Net outperformed VGG19-bn by 40%.
- Mean Average Precision (mAP):
- GPA-Net achieved a 38% improvement in precision over both competing models.
The findings highlight GPA-Net’s robust potential for precise identification, even in challenging forensic scenarios.
Applications in Criminal Investigations
Hand biometrics offer promising applications in forensic investigations, especially for cases involving sexual assaults or crimes where other biometric markers are absent. The researchers emphasize that hands, as primary biometric identifiers, could provide critical evidence through vein pattern recognition and other distinctive traits.
Conclusion
The study by Lancaster University underscores the transformative potential of hand biometrics in forensics. The GPA-Net framework, with its innovative approach to dataset segregation and superior performance metrics, marks a significant advancement in biometric identification. This research paves the way for more reliable, technology-driven solutions to aid forensic investigations and bolster security systems.
More Read
Paper Reference
Baisa, Nathanael L. & Jiang, Zheheng & Vyas, Ritesh & Williams, Bryan & Rahmani, Hossein & Angelov, Plamen & Black, Sue. (2021). Hand-Based Person Identification using Global and Part-Aware Deep Feature Representation Learning. arXiv:2101.05260
