Patching Up Finger Vein Recognition: An Unsupervised Approach using Local Information for Robustness
Tuğçe Arican is a PhD student in the Department of Datamanagement & Biometrics. Promotors are prof.dr.ir. R.N.J. Veldhuis and dr.ir. L.J. Spreeuwers from the Faculty of Electrical Engineering, Mathematics and Computer Science.
In the era where security and privacy are critical, finger vein recognition offers distinct advantages over traditional biometric recognition methods, such as face or fingerprint. Faces and fingerprints are vulnerable to theft and forgery, as faces can be captured in public easily, and fingerprints can be left on surfaces, making them susceptible to unauthorised collection. In contrast, vein patterns are located under the skin, therefore leave no traces and require specialised equipment to be captured. These characteristics position finger vein recognition as a promising technology for applications demanding high level of security and privacy.
Despite the advantages of finger vein recognition, it faces several key research challenges such as limited labelled data, generalisation of learned representations, and lack of tailored image quality assessment methods. Finger vein datasets typically include only a few dozen to a few hundred subjects. Deep learning is a popular approach for finger vein recognition that commonly uses transfer learning, adapting pre-trained models on larger datasets of natural images, for effective learning from limited finger vein datasets. Although this approach achieves state-of-the-art performance on single datasets, when these models are evaluated on new finger vein datasets, their performance significantly decreases. The drop in performance becomes more pronounced when the differences between the training and evaluation sets are larger, indicating the lack of generalisation across datasets with deep learning models. This limitation raises concerns that the learned characteristics from natural images may not sufficiently capture finger vein image characteristics. Additionally, the lack of generalisation limits application of finger vein images in real-world scenarios where multiple devices are involved. Finally, adequate quality assessment for biometric samples includes both image quality and utility aspects, ensuring only suitable images are used for recognition. While image quality defines generic image quality aspects such as sharpness, contrast, and noise, utility evaluates how well a sample supports reliable recognition. Current literature on finger vein image quality often focuses on image quality while overlooking the utility aspect. This can potentially result in misleading quality estimations for finger vein images, posing a risk to reliability and security of the recognition system.
This thesis addresses these challenges by exploring alternative approaches to improve the reliability and robustness of finger vein recognition while providing a deeper understanding of finger vein image characteristics and their influence on recognition and quality estimation.
To address the challenge of learning from limited labelled data, we propose an unsupervised learning approach using an auto-encoder. Auto-encoders can learn robust representations of finger vein images without relying on large datasets. Our auto-encoder captures inherent characteristics of finger vein images, showing promise as an alternative approach for learning from limited data. However, this approach struggles with learning sufficiently discriminative finger vein representations due to the dominance of finger background patterns over vein structures.
Building on the findings of learning from limited data, we modify our auto-encoder into a patch-based approach to address both the challenges learning from limited data and generalisation of learned representations. By dividing finger vein images into patches, we effectively increase the training set size and its variance significantly, while reducing the influence of varying background patterns, enabling the auto-encoder to focus on learning stable vein patterns. Our patch-based auto-encoder not only outperforms the state-of-the-art including a Convolutional Neural Network but also shows consistent performance across various datasets, including unseen ones, without requiring fine-tuning.
Cross-device finger vein recognition is a practical scenario where multiple devices are used, requiring high generalisation of the representations across various device characteristics. To investigate this scenario, we introduce a novel cross-device finger vein dataset including finger vein images of 58 subjects captured by six different acquisition devices. Our findings indicate that device characteristics introduce significant variances in the captured finger vein images, and therefore in the recognition performance. Despite this variability, our patch-based auto-encoder outperforms both a traditional and a deep learning method in cross-device comparisons, indicating its potential for real-world applications. Furthermore, these findings suggest that, besides achieving generalisation with recognition methods, defining and implementing standards for finger vein image acquisition and data exchange is crucial to ensure reliable cross-device comparisons and interoperability for finger vein recognition.
Regarding the quality assessment of finger vein images, we investigate the characteristics that contribute to finger vein image quality, with a focus on their utility for recognition. We propose a patch-based quality assessment approach that measures the visibility and diversity of vein patterns at the local level. By considering these characteristics, our approach aims to ensure that quality estimation accounts for the utility of images in recognition. Evaluations on various dataset shows that this approach is both robust and effective at detecting low-quality images which are not suitable for recognition, showing a potential to achieve reliable and secure finger vein recognition.
The contributions of this thesis are presented as reviewed and published papers, demonstrating the relevance of our work to robust and reliable finger vein recognition. This thesis provides valuable insights into the robustness and reliability of finger vein recognition, offering potential directions for future research improving applicability of finger vein recognition for real-world scenarios.
