Improving 3D Finger Traits Recognition via Generalizable Neural Rendering

Engineering Electrical Engineering Information Science Materials Science Electronic Information Artificial Intelligence Computer Science Bioengineering Life Sciences
3D fingerprint recognition neural rendering finger biometrics 3D biometric traits multi-modal biometrics

Summary of FingerNeRF-Based 3D Finger Biometrics Research Report

Background and Research Significance

With the advancement of biometric technologies, three-dimensional (3D) biometrics have become a promising research direction due to their higher accuracy, robust anti-spoofing capabilities, and resistance to variations in capture angles. Among these, 3D finger biometrics has garnered significant attention in both academia and industry due to its widely used and easily captured biometric traits (e.g., fingerprints, finger veins, and finger knuckles). However, existing 3D biometric methods often rely on explicit 3D reconstruction techniques, which face two primary challenges:

  1. Information Loss: Details are inevitably lost during the explicit reconstruction process, directly impacting recognition performance.
  2. Hardware-Algorithm Coupling: Reconstruction algorithms are often tied to specific hardware, limiting adaptability to different modalities or devices.

To address these issues, researchers proposed FingerNeRF, a method based on implicit Neural Radiance Fields (NeRF), eliminating the need for explicit 3D model reconstruction and enabling direct feature extraction and recognition through learnable neural networks.

Paper Source

This study, titled “Improving 3D Finger Traits Recognition via Generalizable Neural Rendering”, was conducted by researchers Hongbin Xu and colleagues from South China University of Technology. It was published in the “International Journal of Computer Vision”, submitted on September 15, 2023, and accepted on September 14, 2024.

Research Methods and Workflow

The researchers proposed a novel implicit NeRF-based method, FingerNeRF, for solving the 3D finger biometric recognition problem. The main techniques and workflow include:

1. Problem Modeling

Unlike traditional explicit methods requiring 3D model reconstruction, FingerNeRF uses implicit modeling to capture 3D features directly from images. This approach leverages neural rendering, avoiding information loss during reconstruction.

2. Network Architecture and Key Modules

FingerNeRF incorporates the following innovative components during neural rendering: - Trait Guided Transformer (TGT): Enhances cross-view feature alignment guided by traits like fingerprints or veins. - Depth Distillation Loss (DD-Loss): Distills coarse geometric priors from large monocular depth estimation models (e.g., MiDaS) to refine the neural network’s depth map predictions. - Trait Guided Rendering Loss (TG-Loss): Uses feature maps of fingerprints or veins to weight rendering losses, emphasizing key areas for supervision.

3. Dataset Collection and Processing

The research team collected two new datasets: - SCUT-Finger-3D: Includes multi-view images of fingers with clear fingerprint features. - SCUT-FingerVein-3D: Comprises multi-view infrared finger vein images. Additionally, they utilized the publicly available UNSW-3D dataset for evaluation.

Experimental Results and Findings

1. Comparison with Explicit 3D Methods

Compared to traditional explicit reconstruction methods (e.g., COLMAP-based point cloud reconstruction), FingerNeRF demonstrated: - High-quality 3D feature extraction with minimal input views. - Significantly improved recognition performance. For instance, on the SCUT-Finger-3D dataset, the Equal Error Rate (EER) dropped from 35%-40% for explicit methods to 22.60% for FingerNeRF.

2. Adaptability to Different Modalities

FingerNeRF achieved excellent performance across modalities. For example, on the SCUT-FingerVein-3D dataset, FingerNeRF reduced the EER to 16.98%, outperforming explicit methods by 13.37%.

3. Generalization and Scalability

FingerNeRF demonstrated superior generalization across datasets with varying modalities, consistently outperforming existing methods in metrics such as PSNR, SSIM, and LPIPS.

Highlights and Contributions

  1. Method Innovation: Introduced the first implicit NeRF-based framework for 3D finger biometrics, addressing limitations of explicit reconstruction methods.
  2. Multi-Modal Support: Achieved robust performance on different modalities, such as fingerprints and finger veins.
  3. New Datasets: Published two novel multi-view 3D finger biometric datasets, addressing a gap in the research field.
  4. End-to-End Framework: Enabled end-to-end learning for multi-view input to 3D feature extraction via neural rendering.

Research Value and Future Directions

FingerNeRF offers a new perspective for 3D finger biometrics, demonstrating the potential of implicit modeling in biometric recognition. Future directions include extending this method to other biometric traits (e.g., irises, faces), optimizing training processes, and expanding training data diversity. Such improvements could further enhance FingerNeRF’s performance and applicability.