TL;DR
This paper introduces VeinAtnNet, a novel, lightweight convolutional neural network with a multi-head attention mechanism for fingervein verification. It addresses the limitations of existing methods by proposing a new, larger fingervein dataset and demonstrating superior accuracy and efficiency compared to state-of-the-art techniques.
What Problem Does This Paper Address?
Biometric verification systems are crucial for secure access control, and fingervein biometrics offer a compelling solution due to their internal nature, making them resistant to spoofing and external attacks. However, existing fingervein verification methods, particularly those relying on deep learning, face challenges:
- Limited Data: Many current deep learning models are trained and evaluated on small datasets with very few samples per individual, leading to overfitting and hindering their generalization capabilities.
- Inconsistent Evaluation: A lack of standardized evaluation protocols across different studies makes it difficult to reliably compare the performance of various methods.
- Computational Cost: While deep learning models can achieve high accuracy, some can be computationally expensive, limiting their deployment in resource-constrained environments.
This paper aims to overcome these limitations by introducing a more robust and efficient fingervein verification system.
Methods
The core of this work is the VeinAtnNet architecture, a novel convolutional neural network designed for efficiency and accuracy. Key aspects include:
- Convolutional Multi-Head Attention (MHA): VeinAtnNet leverages a combination of convolutional layers for initial feature extraction and a Multi-Head Self-Attention (MSA) mechanism. The MSA allows the network to focus on the most discriminative parts of the fingervein patterns.
- Lightweight Design: The network is intentionally designed to be lightweight, featuring three Consecutive Convolution Layers (CCL) and the MSA module. This aims to reduce the number of learnable parameters while maintaining high performance.
- Dual Input Processing: VeinAtnNet processes both normal (unenhanced) and enhanced fingervein images independently. The enhancement is performed using Contrast Limited Adaptive Histogram Equalization (CLAHE), known for its effectiveness in fingervein imaging.
- Score-Level Fusion: The verification scores generated from the normal and enhanced image branches are fused using a simple sum rule to produce the final verification decision. This approach aims to combine the complementary information from both image types.
- Adaptive Moment Estimation (ADAM) Optimization: The model is trained using the ADAM optimizer with a cross-entropy loss function.
Datasets
The paper introduces a new fingervein dataset and utilizes two publicly available ones:
- FV-300 Dataset (New): This is a newly constructed dataset comprising 300 unique fingervein patterns from 75 data subjects. For each unique fingervein, 92 samples were captured across multiple sessions (1-4 days apart), resulting in a total of 27,600 images. This dataset is made publicly available for research.
- FV-USM Dataset: A publicly available dataset used for evaluation.
- FV-PolyU Dataset: Another publicly available dataset used for evaluation.
The FV-300 dataset is notable for its larger number of samples per identity compared to many existing datasets, addressing the "limited data" problem.
Key Insights and Ideas
- Multi-Head Attention for Discriminative Features: The integration of Multi-Head Attention (MHA) proves crucial for extracting highly discriminative features from fingervein patterns, leading to improved verification accuracy.
- Enhancement Benefits: Processing both normal and enhanced fingervein images and fusing their scores yields better performance, especially at lower False Match Rates (FMR). Enhanced images seem to highlight vein patterns and minutiae more effectively.
- Lightweight Architecture Sufficiency: The study demonstrates that a lightweight CNN architecture with MHA can achieve state-of-the-art performance, challenging the notion that deeper and more complex models are always necessary.
- Dataset Size Matters: The performance improvements observed on the newly introduced FV-300 dataset highlight the impact of having larger, more diverse datasets for training and evaluation.
Key Results
- Superior Performance on FV-300: The proposed VeinAtnNet achieved an outstanding Equal Error Rate (EER) of 0.54% and a True Accept Rate (TAR) of 90.36% at an FMR of 0.01% on the FV-300 dataset.
- Competitive Performance on Public Datasets: The method also showed strong results on publicly available datasets, achieving an EER of 15.35% on FV-USM and 5.52% on FV-PolyU.
- Efficiency: The VeinAtnNet is significantly lighter than some existing SOTA methods, boasting only 58.2K learnable parameters compared to millions in some other approaches.
- Ablation Study: An ablation study confirmed that adding more convolutional layers alongside the MHA module generally improves performance, with the proposed three-layer CNN + MHA configuration striking a good balance.
- LIME Explainability: Qualitative analysis using LIME showed that enhanced fingervein images allow the model to focus on smaller, more discriminative regions (like minutiae) for verification.
Limitations and Future Work
- Generalization of Deep Learning Models: The paper notes that deep learning techniques can struggle to generalize to new datasets due to limited fine-tuning data, though VeinAtnNet performs competitively.
- Further Investigation of Convolutional Layers: While the study found that three convolutional layers were optimal, the authors suggest further exploring deeper architectures with MHA for potentially even better performance.
- Exploring Other Fusion Strategies: The current score-level fusion is a simple sum rule. Future work could explore more sophisticated fusion techniques.
- Real-World Deployment: While the dataset is publicly available, further testing and optimization would be needed for deployment in real-world, diverse scenarios.
Summary updated