CGAN-DA: A Cross-Modality Domain Adaptation Model for Hand-Vein Biometric-based Authentication

Shuqiang Yang, Yiquan Wu, Xin Jin, Mounîm A. El‐Yacoubi, Huafeng QinJOURNAL 2022 · DOI

TL;DR

This paper introduces CGAN-DA, a novel unsupervised domain adaptation model for hand-vein biometric authentication. By leveraging retinal vein images with abundant annotations, CGAN-DA effectively bridges the domain gap to palm-vein images, achieving state-of-the-art verification accuracy without requiring any labeled palm-vein data.

What Problem Does This Paper Address?

Biometric authentication systems, like those using palm-vein patterns, are crucial for security. While deep learning models, particularly CNNs, excel at learning robust features for accurate recognition, they typically require large, meticulously annotated datasets. For vein biometrics, obtaining such pixel-level annotations for palm-vein images is prohibitively expensive and time-consuming. Existing methods often rely on handcrafted features or less robust pseudo-annotations, limiting their performance. This paper addresses the challenge of achieving high-accuracy palm-vein recognition in an unsupervised manner, by transferring knowledge from a related domain (retinal veins) that does have readily available annotations.

Methods

The core of the proposed solution is the CGAN-DA (CycleGAN-based Domain Adaptation) model, which combines two key adaptation strategies:

  • Image Adaptation: This stage uses a CycleGAN architecture to transform retinal vein images (source domain) into synthetic palm-vein images (target domain appearance). This process aims to align the visual characteristics of the two domains, reducing the domain shift at the input level. It employs adversarial losses and a cycle-consistency loss to ensure that the transformations are realistic and reversible.
  • Feature Adaptation: Following image adaptation, a synergistic learning framework is used to extract domain-invariant features. This involves a shared encoder and two classifiers. An additional GAN-based discriminator is introduced to further minimize the remaining domain shift between the synthesized target images and real palm-vein images. This stage focuses on learning features that are robust and discriminative, regardless of the original domain.

The model integrates these two adaptation strategies to jointly learn image appearance and feature representations, thereby mitigating the domain gap effectively.

Datasets

  • DRIVE Dataset: Used for retinal images and their associated manual vasculature annotations. This dataset serves as the source domain with rich ground truth.
  • CASIA Multi-Spectral Palm-print Image Dataset: Used for palm-vein images. For this work, images captured under the 850 nm wavelength were utilized. This dataset serves as the target domain, and importantly, the model operates without requiring pixel-level annotations from this dataset.

The paper describes a preprocessing step where retinal images are cropped and scaled to match the resolution of palm-vein images, creating synthetic training data for the image adaptation phase.

Key Insights and Ideas

  • Leveraging Related Domains for Unsupervised Learning: The paper demonstrates the power of cross-modality domain adaptation by successfully transferring knowledge from retinal vein images to palm-vein authentication, bypassing the need for expensive palm-vein annotations.
  • Synergistic Image and Feature Adaptation: Combining both image-level style transfer and feature-level domain invariance is crucial for effectively bridging the significant visual and structural differences between retinal and palm-vein patterns.
  • CycleGAN for Realistic Synthesis: The use of CycleGAN for generating synthetic palm-vein images from retinal ones is key to providing a realistic target domain appearance for subsequent feature learning.
  • Domain-Invariant Features for Robustness: The feature adaptation stage is critical for ensuring that the learned representations are not overly sensitive to the domain shift, leading to improved generalization and accuracy.

Key Results

  • Lowest Equal Error Rate (EER): The proposed CGAN-DA model achieved the lowest EER of 0.52% on the CASIA dataset, significantly outperforming existing methods.
  • Outperforms Benchmarking Approaches: Compared to handcrafted methods (e.g., Repeated line tracking, Gabor filters) and a CNN-based approach, CGAN-DA demonstrated superior verification accuracy.
  • CNN Performance: The CNN-based approach achieved an EER of 0.74%, which is still considerably higher than CGAN-DA.
  • Qualitative Improvements: Visual assessments showed that CGAN-DA extracts vein patterns that are more connected, smooth, and less noisy compared to other methods.

Limitations and Future Work

  • Image Preprocessing: The authors acknowledge that the patch division and scaling of retinal images are important steps, as vascular thickness can differ between eyes and hands.
  • Threshold Selection: The current model uses a fixed threshold (0.5) for segmenting vein patterns. The authors suggest that this threshold might not be optimal and propose exploring search-based approaches to find a more objective threshold that directly relates to verification error reduction.
  • Future Directions:
    • Using the proposed model to generate ground-truth vein images for training existing deep learning models.
  • Conducting further experiments on other large vein datasets to validate the model's efficiency.
  • Investigating alternative domain adaptation approaches for vein pattern segmentation.

Summary updated

Abstract

Palm-vein recognition has been the focus of large research efforts over the last years. However, despite the effectiveness of deep learning models, in particular Convolutional Neural Networks (CNNs), in automatically learning robust feature representations, thereby obtaining good accuracy, such good performance is usually obtained at the expense of annotating a large training dataset. Labeling vein images, however, is an expensive and tedious process. Although handcrafted schemes for data augmentation usually increase slightly performance, they are unable to cover complex variations inherently characterizing such images. To overcome this issue, we propose a new unsupervised domain adaptation model, called CycleGAN-based domain adaptation (CGAN-DA), that extracts discriminant representation from the palmvein images, without requiring any image labeling. Our CGAN-DA models allows a conjoint adaptation, at the image and feature levels. Specifically, in order to enhance the extracted features’ domain-invariance, image appearance is transformed across two domains, palm-vein domain and retinal domain. We employ several adversarial losses namely a segmentation loss and a cycle consistence loss to train our model without any annotation from the target domain (palm-vein images). Our experiments on the public CASIA palm-vein dataset demonstrates that our models significantly outperforms the s tart of the art in terms of verification accuracy.
BibTeX
@article{yang2022cgan,
  author = {Shuqiang Yang and Yiquan Wu and Xin Jin and Mounîm A. El‐Yacoubi and Huafeng Qin},
  title = {{CGAN-DA: A Cross-Modality Domain Adaptation Model for Hand-Vein Biometric-based Authentication}},
  journal = {JOURNAL},
  year = {2022},
  doi = {10.61702/yrmc5182},
}
Figure