Fingerprint recognition is a long-standing and important topic in computer vision and pattern recognition research, supported by its diverse applications in real-world scenarios such as access control, consumer products, law enforcement, forensics, national identity, and border security. Recent advances in deep learning have greatly enhanced fingerprint recognition accuracy and efficiency alongside traditional hand-crafted fingerprint recognition methods, particularly in controlled settings. While state-of-the-art fingerprint recognition methods excel in controlled scenarios, like rolled fingerprint recognition, their performance tends to drop in uncontrolled settings, such as latent and contactless fingerprint recognition. These scenarios are often characterized by extreme degradations and image variations in the captured images. This performance drop is due to the inability of fingerprint embeddings (feature vectors obtained via deep networks) to generalize across variations in the captured fingerprint images between controlled and uncontrolled settings. The challenges in the generalization of fingerprint embeddings, from controlled to uncontrolled settings, encompass issues such as insufficient labeled data, varying domain characteristics (often referred to as “domain gap”), and the misalignment of fingerprint features due to information loss. This dissertation proposes a series of methods aimed at addressing these challenges in various unconstrained fingerprint recognition scenarios. We begin in chapter 2 with an examination of cross-sensor and cross-material presentation attack detection (PAD), where the sensing mechanism and encountered presentation attack instruments (PA) may be unknown. We present methods to augment the given training data to include a wider diversity of possible domain characteristics, while simultaneously encouraging the learning of domain-invariant representations. Next, we turn our attention in chapter 3 to the challenging scenario of contact to contactless fingerprint matching, where misaligned fingerprint features due to differences in contrast, perspective differences, and non-linear distortions are corrected via a series of deep learning-based preprocessing techniques to minimize the domain gap between contact and corresponding contactless fingerprint images. In chapter 4, we aim to improve the sensor-interoperability of fingerprint recognition by leveraging a diversity of deep learning representations, integrating convolutional neural network and attention-based vision transformer architectures into a single, multimodel embedding. Similarly, in chapter 5, we further improve the robustness and universality of fingerprint representations by fusing multiple local and global embeddings and demonstrate a marked improvement in latent to rolled fingerprint recognition performance, both in terms of accuracy and efficiency. Next, chapter 6 presents a method for synthetic fingerprint generation, capable of mimicking the distribution of real (i.e., bona fide) and PA (i.e., spoof) fingerprint images, to alleviate the lack of publicly available data for building robust fingerprint presentation attack detection algorithms. Finally, in chapter 7 we extend our fingerprint generation capabilities toward generating universal fingerprints of any fingerprint class, acquisition type, sensor domain, and quality, all to improve fingerprint recognition training and generalization performance across diverse scenarios.
Read
