dc.description.abstract
Fingerprint recognition is a widely used biometric modality because of its uniqueness and persistence of friction ridge patterns, which provide a reliable means of verifying identity. It is a key technology in applications ranging from law enforcement and border control to securing personal devices and financial transactions. While traditionally acquired through direct contact, contactless methods offer advantages in hygiene and user convenience. However, practical deployment is challenged by the unconstrained nature of the acquisition process, which introduces variations in finger pose, illumination, and scale, while still requiring to be interoperability with large, legacy contact-based fingerprint databases. This thesis presents a set of algorithms and methodologies to address these problems across the recognition pipeline, from image capture to secure template comparison. The contributions include solutions for image normalization, data assurance, robust validation, and secure deployment, which improve the accuracy, reliability, and privacy of contactless fingerprint systems.A necessary step in any contactless pipeline is the accurate segmentation of the fingertip from its background, which is often complex and variable. This work proposes three novel deep learning architectures for this task. The first is a custom U-Net-based model for pre-cropped single-finger images that outperforms existing segmentation models [217]. This was improved with FingerUNeSt++, which combines a ResNeSt encoder with a UNet++-like decoder, achieving a mean Intersection-over-Union (mIoU) of 99% on the test set. The third model, TipSegNet, removes the need for a separate finger detection step by segmenting and labeling all four fingertips directly from a whole-hand image. Using a ResNeXt-101 backbone with a Feature Pyramid Network (FPN) to handle multi-scale objects, TipSegNet obtains an mIoU of 99% and an accuracy of 100%.Interoperability with contact-based systems requires correcting the geometric distortions in contactless captures. This research developed a processing pipeline to correct for in-plane (yaw) and out-of-plane (roll) rotations, which then flattens the fingertip texture using parametric unwarping models. The pipeline uses the segmentation mask for yaw correction and an elliptical finger model with the detected core for roll correction. On an operational dataset, a finger-wise optimized application of the pipeline reduced the Equal Error Rate (EER) for contactless-to-contact-based comparison by a relative 36.9% (from 1.57% to 0.99%). A large-scale empirical analysis of the fingerprint core’s position across over 40,000 samples showed that its location is not geometrically centered, with systematic, modality-induced biases and a natural variability of 6-12% of the finger’s width. This study quantifies a limit on the accuracy of alignment methods that rely on the core and identifies the Non-Central Fischer (NCF) distribution as the best-fitting statistical model for its position for most fingers, including a finger-dependent analysis.The quality of a captured sample affects recognition performance. This work addresses the absence of dedicated quality metrics for mobile contactless fingerprints by adapting the established NFIQ 2 framework, resulting in MCLFIQ. By retraining the NFIQ 2 random forest classifier on modality-specific synthetic data, MCLFIQ shows improved performance in predicting the utility of contactless samples compared to the original NFIQ 2.2 and other baselines. The new model prioritizes features related to image sharpness and local ridge clarity, which are key quality factors in mobile captures. Additionally, this thesis introduces a self-supervised framework to detect structural artifacts from fingerprint mosaicking, a problem not handled by standard quality metrics. A deep learning model was trained on programmatically generated artifacts to detect these defects without manual annotation. The resulting detector is accurate (IoU > 0.9), robust to other quality defects, and generalizes across different fingerprint modalities (contactless, rolled, slap).The validation of these algorithms requires standardized testing methods. This research investigated the manufacturing of physical synthetic fingerprint phantoms, evaluating various fabrication techniques (direct laser engraving, CNC machining, 3D printing of molds) and materials (silicone, gelatin, elastomer). The work shows that high-fidelity, stable, and realistic 3D phantoms can be produced, particularly using silicone cast in high-precision molds. These phantoms replicate a known ground-truth pattern and provide a repeatable method for evaluating sensor and algorithm performance.For privacy-sensitive applications like watchlist checks, this thesis presents a practical privacy-preserving comparison solution. The minutiae-based comparison algorithm SourceAFIS was adapted for execution within a Multiparty Computation (MPC) framework. Through algorithmic refactoring and the use of plaintext computations for non-sensitive data, the implementation achieves a comparison time of approximately 17 seconds while maintaining high recognition accuracy comparable to its plaintext counterpart. This makes secure remote fingerprint comparison practical for real-world scenarios.Finally, this thesis shows that the challenges accompanying a complete, end-to-end pipeline for secure contactless fingerprint recognition can be solved. By integrating the developed solutions, one can go from raw images captured to standardized and interoperable biometric templates. The relevance of this integrated approach is its ability to enable high-security applications on contactless systems, such as an agent performing a watchlist check at a border using a standard smartphone.
en
