Adaptive Learning Gabor Filter for Finger-Vein Recognition

Yakun Zhang, Weijun Li, Liping Zhang, Xin Ning, Linjun Sun, Yaxuan LuIEEE Access 2019 · OpenReview

Abstract

Presently, finger-vein recognition is a new research direction in the field of biometric recognition. The Gabor filter has been extensively used for finger-vein recognition; however, its parameters are difficult to adjust. To solve this problem, an adaptive-learning Gabor filter is presented herein. We combine convolutional neural networks with a Gabor filter to calculate the gradient of the Gabor-filter parameters, based on the objective function, and to then optimize its parameters via back-propagation. The parameter $\theta$ of Gabor filter can be trained to the same angle as the vein texture of finger vein image. The parameter $\sigma$ of Gabor filter has a certain relation with $\lambda$, and the parameter $\lambda$ of Gabor filter can converge to the optimal value well. Using this method, we not only select appropriate and effective Gabor filter parameters to design the filter banks, we also consider the relationship between those parameters. Finally, we perform experiments on four public finger-vein datasets. Experimental results demonstrate that our method outperforms state-of-the-art methods in finger-vein classification.

Introduction

Finger-vein recognition is a promising biometric modality due to its high security—vein patterns are internal and difficult to forge. However, the low contrast and uneven illumination of raw finger-vein images make feature extraction challenging. The Gabor filter is a traditional tool for this task, valued for its ability to capture multi-scale and multi-orientation textures. Despite its effectiveness, the performance of Gabor filters is highly sensitive to hyper-parameters like orientation ($\theta$), wavelength ($\lambda$), and bandwidth ($\sigma$), which are typically tuned manually or through exhaustive search.

Methodology: Adaptive Gabor CNN

This research proposes a hybrid architecture that integrates the mathematical rigor of Gabor filters with the learning power of Convolutional Neural Networks (CNNs). Instead of learning arbitrary weights, the first layer of the network is constrained to follow the 2D Gabor function. By deriving partial derivatives for these parameters, the model can optimize the filter bank directly through backpropagation.

Figure 1: A bank of Gabor filters at various scales and orientations, representing the initial or learned states of the adaptive layer.

The architecture consists of a Gabor convolutional layer with 16 channels, followed by ReLU activation, a fully connected layer for feature dimensionality reduction (128-D), and a softmax loss function for classification.

Key Insights and Visualizations

The primary advantage of this approach is its ability to align filter parameters with the physical characteristics of the input data.

Orientation Alignment

As the network trains, the orientation parameter $\theta$ converges to match the dominant directions of the vein textures. This adaptive alignment ensures that the filters capture the most salient features regardless of finger orientation.

Figure 2: (a) Original ROI image and (b)-(i) the resulting feature maps after applying Gabor filters at different orientations ($\theta$).

Parameter Coupling

The study found that the bandwidth $\sigma$ and wavelength $\lambda$ maintain a stable ratio (typically 0.27–0.39) after training, which corresponds to the optimal spatial-frequency bandwidth for texture analysis. This suggests the network naturally discovers signal processing principles.

Experimental Results

The method was evaluated on four public datasets: MMCBNU_6000, FV-USM, SDUMLA, and HKPU. The adaptive Gabor filter consistently achieved lower Equal Error Rates (EER) than traditional handcrafted methods and standard CNNs.

  • MMCBNU_6000: Achieved a 0.11% EER, significantly lower than the 0.42% of the previous best method.
  • HKPU: Achieved 1.67% EER, nearly halving the error rate of some deep representation-based methods (3.02%).

The model's success is attributed to its strong inductive bias; by using Gabor filters as a structural prior, it requires less training data to achieve high generalization compared to "black-box" CNNs.

Conclusion

The Adaptive Learning Gabor Filter effectively automates parameter tuning for finger-vein recognition. By embedding domain-specific knowledge into a deep learning framework, the researchers achieved state-of-the-art performance with high interpretability. Future work will explore deeper adaptive architectures and the fusion of soft biometric traits.

Summary updated

Abstract

Presently, finger-vein recognition is a new research direction in the field of biometric recognition. The Gabor filter has been extensively used for finger-vein recognition; however, its parameters are difficult to adjust. To solve this problem, an adaptive-learning Gabor filter is presented herein. We combine convolutional neural networks with a Gabor filter to calculate the gradient of the Gabor-filter parameters, based on the objective function, and to then optimize its parameters via back-propagation. The parameter θ of Gabor filter can be trained to the same angle as the vein texture of finger vein image. The parameter σ of Gabor filter has a certain relation with λ, and the parameter λ of Gabor filter can converge to the optimal value well. Using this method, we not only select appropriate and effective Gabor filter parameters to design the filter banks, we also consider the relationship between those parameters. Finally, we perform experiments on four public finger-vein datasets. Experimental results demonstrate that our method outperforms state-of-the-art methods in finger-vein classification.
BibTeX
@inproceedings{zhang2019adaptive,
  author = {Yakun Zhang and Weijun Li and Liping Zhang and Xin Ning and Linjun Sun and Yaxuan Lu},
  title = {{Adaptive Learning Gabor Filter for Finger-Vein Recognition}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=FC1piEIeyu},
}
Figure