AiRDUNet: All-In-One Residual Dense U-Net with Automated Humidity Quality Assessment for Partial Wet Fingerprint Restoration and Recognition

The recognition performance on smartphones equipped with side-mounted sensors often suffers from a limited size in the fingerprint images with deteriorated quality. Although most studies leverage deep learning to enhance fingerprint quality, this paper investigates the impact of fingerprint humidity on image quality and the restoration of partially wet fingerprints. We introduce a Wet Fingerprint Quality Automated Assessment (WFQA) framework that uses six humidity-related indicators, an automated labeling system, and a KNN classifier to define fingerprint humidity levels.By integrating WFQA with CycleGAN, we generate synthetic wet fingerprints for data augmentation under various humidity conditions. Additionally, we developed the All-in-One Residual Dense-UNet (AiRDUNet) for restoration across different humidity levels, using residual dense blocks and a recognition loss function to prevent over-restoration and preserve recognition features. AiRDUNet effectively restores wet fingerprints while maintaining recognition accuracy, outperforming existing methods on real-world test sets with an FRR of 11.89%. AiRDUNet shows improvements of 18.84% and 34.23% over DenseUNet and Residual-Mnet, respectively, and surpasses PGT-Net and FPN-ResUNet by 12.78% and 10.26%.