Residual M-net with Frequency-Domain Loss Function for Latent Fingerprint Enhancement

Fingerprints generally deposited involuntarily on a surface and that usually require special chemical procedures to become visible are called latent fingerprints. These fingerprints have a relevant role in identifying individuals in forensic applications. In many cases, latent fingerprint images are inherently noisy, exhibit perturbations produced by their forming mechanism, and convey limited information, which requires enhancement procedures to improve the identification result. However, most existing methods successfully used to enhance fingerprints collected in controlled environments do not perform as well on latent fingerprints. In this research, we propose a latent fingerprint enhancement method based on a deep learning model that consists of a residual encoder-decoder architecture that optimizes a frequency-domain loss function during training. In addition, a procedure for synthetically generating a large set of labeled training data has been developed. Experiments with the MOLF database have shown that our proposed method improved the quality of fingerprints and resulted in better minutiae extraction, thus improving fingerprint identification performance. Compared with other methods published in the literature, the proposed method achieved better accuracies in ranks 1 to 30 according to CMC curves.