Ensemble-Based Bounding Box Regression for Enhanced Knuckle Localization

Area of Science:

• Biometrics
• Computer Vision
• Forensic Science

Background:

• Facial recognition is common, but not always available for identification.
• Knuckle creases on the dorsal hand present a viable alternative biometric trait.
• Accurate localization of knuckle creases is crucial for effective identification.

Purpose of the Study:

• To develop an ensemble approach for accurate knuckle region localization on the dorsal hand.
• To improve the efficacy of hand-based biometrics for forensic identification.
• To evaluate the generalizability of the proposed localization method across diverse datasets.

Main Methods:

• Utilized an ensemble of multiple object detector frameworks for knuckle region localization.
• Tested the approach on two large-scale public hand databases and a novel proprietary dataset.
• Developed novel performance metrics to assess the accuracy of knuckle region detection against ground truth.

Main Results:

• The ensemble approach demonstrated superior performance in localizing knuckle regions compared to individual detectors.
• The method showed effectiveness across varying backgrounds and finger positions.
• Cross-dataset evaluation confirmed the generalizability and robustness of the proposed approach.

Conclusions:

• Ensemble object detection provides a comprehensive and accurate method for dorsal hand knuckle localization.
• This technique enhances the potential of hand-based biometrics in forensic applications.
• The proposed approach offers a reliable solution for identifying individuals when other biometrics fail.