Abstract
This paper proposes a hierarchical transfer learning strategy to address the "small-data" challenge in hand vein recognition. By bridging the domain gap between large-scale visible-light face datasets and small-scale near-infrared (NIR) vein datasets through an intermediate NIR face domain, the authors develop a bimodal system for identity and gender recognition. The approach achieves a 91.6% gender classification accuracy and a 0.013% Equal Error Rate (EER) for identity recognition, significantly outperforming traditional hand-crafted features and direct transfer learning methods.
Introduction
Deep Convolutional Neural Networks (DCNNs) have revolutionized biometric recognition, particularly in face recognition where datasets like VGG-Face contain millions of images. However, hand vein recognition faces a significant "small-data" hurdle. Furthermore, a substantial domain gap exists between standard RGB face images and near-infrared (NIR) vein images. Direct fine-tuning of a face-trained model on a small vein dataset often leads to overfitting or poor feature extraction.
This research introduces a "coarse-to-fine" hierarchical transfer learning strategy to bridge this gap. The goal is to create a bimodal system capable of identifying both an individual's identity and gender using only hand vein data, even when training data is scarce.
Methods
The core innovation is a Hierarchical Knowledge Transfer procedure. Instead of jumping directly from visible-light faces to NIR veins, the authors use a four-step "general-to-specific" pipeline to gradually adapt the model to the target domain:
- Source Model: The process begins with the VGG-Face model, pre-trained on 2.6 million visible-light face images.
- Intermediate Step (Transitional Model): The model is fine-tuned on the PolyU NIR-face database. This stage teaches the network the characteristics of near-infrared imaging while maintaining familiar facial structures.
- Identity Task (VIM): The resulting model is fine-tuned on a lab-made hand vein database to recognize specific individuals (Vein Identity Model).
- Gender Task (VGM): Finally, the identity model is fine-tuned on the same vein database but labeled for gender (Vein Gender Model). This step exploits the correlation between identity-specific features and gender-specific traits.
For the final classification, features are extracted from the second-to-last fully connected layer (FC7). These features are then classified using a Linear Support Vector Machine (SVM), which provides a robust decision boundary for both identity and gender tasks.
Key Insights
- The "Shared Pattern" Hypothesis: The authors argue that neighboring domains, such as NIR faces and NIR veins, share similar low-level patterns (noise, contrast, texture). By stepping through these domains, the model adapts to the target task without losing the general feature-extraction power gained from the massive source dataset.
- Identity-to-Gender Transfer: Training for identity first, then gender, proved more effective than training for gender from scratch. This suggests that the structural features required to distinguish individuals are a superset of those needed for gender classification.
- Regularization via Hierarchy: The hierarchical approach acts as a form of regularization, guiding the weights toward a useful solution and preventing the overfitting typically seen when training deep networks on small datasets.
Results
The proposed method significantly outperformed traditional hand-crafted feature extractors (such as SIFT and LBP) and direct transfer learning approaches:
- Gender Classification: Achieved 91.60% accuracy. In contrast, directly fine-tuning VGG-Face on veins yielded only 29.56% accuracy.
- Identity Recognition: Achieved an Equal Error Rate (EER) of 0.013%, substantially better than the best hand-crafted method (RootSIFT at 0.105%).
- Robustness: The model demonstrated high resistance to non-uniform illumination and pose variations, common challenges in biometric systems.
Conclusion
The "coarse-to-fine" hierarchical fine-tuning strategy effectively addresses the data scarcity and domain gap issues in vein recognition. While the current study used a lab-made database of 200 subjects, the results indicate a promising path for high-accuracy biometric systems. Future work will involve testing on larger-scale, "in-the-wild" datasets to further validate commercial viability and exploring the application of this strategy to other vision tasks with limited data.
Summary updated