ResUNet: A Lightweight Residual U-Net for Scalable Fingerprint Enhancement

• André Nóbrega UNICAMP
• João Contreras UNICAMP
• Pascual Figueroa Griaule
• Alexandre Falcão UNICAMP

Resumo

Enhancing latent fingerprints is a critical step in forensic biometric pipelines that remains challenging due to severe image degradation and the need for scalable processing. In this work, we introduce ResUNet, a lightweight Residual U-Net architecture for fingerprint enhancement that achieves competitive identification accuracy while remaining highly efficient. Our framework leverages a novel synthetic latent fingerprint generator designed to simulate realistic noise, occlusions, and background patterns for self-supervised learning. By segmenting recoverable regions with FingerNet and excluding unrecoverable areas from supervision, our model avoids hallucination, an issue prevalent in prior methods. We evaluate our method on the NIST SD27 dataset under a closed-set identification protocol using a 20,473-image gallery. Our approach outperforms recent state-of-the-art lightweight models by 6% Rank-1 accuracy while being up to 180× faster than transformer-based and GAN-based solutions. We also release our synthetic generation framework to promote reproducibility and benchmarking. These results show that ResUNet is a scalable and reliable solution for operational AFIS.

Palavras-chave: Degradation, Accuracy, Protocols, Forensics, Noise, Fingerprint recognition, NIST, Transformers, Generators, Synthetic data