Patching up finger vein recognition: An unsupervised approach using local information for robustness

In the era of heightened security and privacy concerns, finger vein recognition presents a significant advantage over traditional biometric methods like facial and fingerprint recognition. Unlike fingerprints and facial features, which can be easily captured or replicated, vein patterns are hidden beneath the skin, making them resistant to theft and forgery. This makes finger vein recognition a promising technology for high-security applications.

However, several challenges hinder its widespread adoption. These include limited labelled data, difficulties in generalising learned representations, and the lack of comprehensive image quality assessment methods. Finger vein datasets are relatively small, making deep learning approaches reliant on transfer learning from natural image datasets. While this technique achieves state-of-the-art performance on individual datasets, it struggles with cross-dataset generalisation, limiting its effectiveness in real-world applications where multiple devices are involved. Additionally, current image quality assessments often focus on general image attributes like sharpness and contrast, overlooking their actual utility in recognition.

To address these challenges, this thesis explores alternative approaches for improving the reliability and robustness of finger vein recognition. An unsupervised auto-encoder model is proposed to learn meaningful representations without requiring large datasets. However, it struggles with distinguishing vein structures due to the dominance of background patterns. To overcome this, a patch-based approach is introduced, dividing images into smaller sections to enhance training variability and reduce background interference. This approach outperforms traditional deep learning models and maintains consistent performance across unseen datasets.

For cross-device recognition, a novel dataset comprising images from six different devices is introduced. The results show that device variations impact recognition performance, but the patch-based auto-encoder achieves superior generalisation compared to traditional and deep learning methods. This highlights the need for standardised acquisition protocols for improved interoperability.

Finally, a patch-based quality assessment method is developed, focusing on vein visibility and diversity to ensure reliable recognition. Evaluations demonstrate its effectiveness in detecting poor-quality samples, enhancing system security.

Overall, this research advances finger vein recognition technology, offering insights into its robustness, generalisation, and quality assessment, with potential for real-world applications.