LightFinger-GAN: lightweight generative enhancement for latent fingerprint recognition

The recognition of latent fingerprints continues to be a difficult task due to the substantial noise, absent ridge patterns, low contrast, and structural abnormalities in latent impressions. These factors restrict the extraction of minutiae and downstream matching. This study proposes LightFinger-GAN , a lightweight generative framework for latent fingerprint enhancement that is connected to a hierarchical minutiae-based matching technique for robust recognition, in order to overcome these constraints. The proposed enhancement model regards latent fingerprint restoration as a restricted fingerprint-to-fingerprint translation problem, with the goal of minimizing irrelevant background anomalies and recovering ridge structures and minutiae-consistent characteristics. A synthetic latent fingerprint generation approach is established to supplement the training process with possible deterioration patterns, thereby relieving the lack of paired training data. The minutiae are extracted after enhancement and encoded by local geometric relationships. A two-stage hierarchical matching scheme is implemented to enhance the discrimination between the genuine and impostor fingerprint pairs. The experiments on public datasets demonstrate that the proposed framework has better enhancement quality and matching performance than the state-of-the-art methods. The proposed hierarchical matcher achieves a minimum Equal Error Rate (EER) of 0.05% on the CASIA dataset, while also reducing false minutiae and recovering a greater number of genuine minutiae. The results show the proposed model is a reliable and computationally efficient solution for latent fingerprint identification in security-critical biometric applications.