Robust 3D finger knuckles biometric identification with hierarchical featureNet architecture

Abstract

A novel biometric identifier known as the 3D finger knuckle pattern provides highly discriminative characteristics for the finger knuckle-based personal identification. This paper addresses the challenge of 3D finger knuckle recognition, aiming to enhance precision and overcome limitations in existing approaches. Leveraging neural network technology, it introduces a novel neural network framework for this purpose. Recent research has made significant progress in 3D finger knuckle recognition, particularly in the areas of matching schemes, feature representations, and specialized deep neural networks. Challenges such as limited training data and dataset heterogeneity are discussed. The proposed 3D hierarchical featureNet (HFN) methodology involves a multi-stage pre-processing process for 3D images, encompassing detection, cropping, smoothing, and hole-filling. Feature similarity is evaluated with nearest neighbor distance ratios, enabling precise recognition. The key contribution of this work is the introduction of a new feature vector that incorporates curvature data, improving the state-of-the-art. The methodology employs statistical distribution analysis for feature similarity and 3D geometry for accurate curvature representation. Overall, this research offers a comprehensive solution for 3D finger knuckle recognition, enhancing accuracy and efficiency through innovative pre-processing, feature extraction, and similarity evaluation methods.