Outpainting Fingerprint Images: Expanding Partial Prints with Generative Models

Abstract

This thesis presents a generative deep learning framework for outpainting of missing regions in partial fingerprint images. The architecture builds on a U-Transformer backbone, incorporating Swin Transformer blocks and trained within a GAN setup using the WGAN-GP loss to ensure stable adversarial learning. The model learns to create plausible extensions of fingerprint images, maintaining ridge structures and generating visually coherent and structurally consistent completions.

Beyond general outpainting, the framework is also adapted to generate artifact fingerprints for data augmentation purposes. Additionally, by modifying the loss functions, the model is adapted to generate structural defects caused by scars. A lightweight U-Net segmentation model is first trained to identify scarred areas in real fingerprint images, and these regions are then used to guide the inpainting model to generate realistic scar patterns within the damaged areas, preserving the surrounding ridge structure.
The scar generation inpainting model achieves strong reconstruction fidelity in small, irregularly damaged regions and opens possibilities for generating synthetic scar patterns, enabling future applications in fingerprint dataset augmentation. Its successful extension to scar inpainting highlights the framework’s versatility for both augmentation with artifacts and broader fingerprint restoration tasks.

The models are evaluated using standard image similarity metrics, ridge continuity, and visual inspection, showing promising results across all use cases.

Popular Abstract

Fingerprint recognition plays a central role in many modern security and identification systems. However, in practice, fingerprint images are often incomplete — whether due to sensor limitations, privacy regulations, or the practical challenges of data collection. This thesis addresses that challenge by developing deep learning models capable of generating plausible extensions to partial fingerprint images, a task referred to as outpainting.
At the core of the proposed solution is a generative deep learning framework based on the U-Transformer architecture. This model combines the strengths of U-Net — known for preserving spatial detail through skip connections — with Swin Transformer blocks that capture long-range dependencies across the image. Training is performed within a Generative Adversarial Network (GAN) setup, using a stabilized learning approach called Wasserstein GAN with Gradient Penalty (WGAN-GP). Together, these components allow the model to generate high-quality, visually coherent and structurally consistent fingerprint completions, extending ridge lines and maintaining realism even when large areas are missing.
Beyond general-purpose outpainting, the same framework is adapted for two specialized fingerprint restoration tasks:

Artifact Outpainting: This variant applies the outpainting model to reconstruct fingerprint regions corrupted by sensor-like artifacts. Unlike the general model, where missing regions are randomly defined, the corrupted areas here are predefined per image. While the same architecture and loss functions are used, the focus shifts toward replacing known defects with plausible fingerprint content — supporting robust biometric systems through improved data augmentation.


Scar Inpainting: This model variant is adapted with minimal modifications to inpaint internal fingerprint defects, such as scars. It learns to restore structured interruptions in the ridge flow while preserving both visual realism and identity consistency. In addition to reconstructing naturally scarred areas, the model can be used to generate realistic synthetic scars, offering a valuable tool for augmenting fingerprint datasets with diverse and challenging samples.

The models are evaluated using standard image similarity metrics, ridge continuity analysis, and fingerprint matching performance. Visual examples demonstrate high-quality reconstructions, especially when partial input data is available. Results also show that outpainting can modestly improve biometric verification accuracy in degraded cases — though the impact varies depending on how much fingerprint data is missing and the matching thresholds used. The results further demonstrate that the outpainting framework can be effectively adapted to scar inpainting.
In conclusion, this work presents a versatile generative framework for reconstructing and augmenting fingerprint images. Whether the goal is to complete arbitrary missing regions or enrich datasets with synthetic examples, the proposed models demonstrate promising results across all use cases. The framework offers a practical contribution to improving the reliability, robustness, and realism of modern biometric systems.

• author: Weiber, Joakim
• supervisor: Alexandros Sopasakis
• organization: Mathematics (Faculty of Engineering)
• course: FMAM05 20251
• year: 2025
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Outpainting, Deep Generative Models
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMA-3596-2025
• ISSN: 1404-6342
• other publication id: 2025:E65
• language: English
• id: 9207040
• date added to LUP: 2025-07-01 10:38:28
• date last changed: 2025-07-01 10:38:28