Latent Fingerprint Enhancement Based on DenseUNet

Abstract

The image quality of latent fingerprints is usually poor with unclear ridge structure and various overlapping patterns. Enhancement is an important processing step to reduce the noise, recover the corrupted regions and improve the clarity of ridge structure for more accurate fingerprint recognition. Existing fingerprint enhancement methods cannot achieve good performance for latent fingerprints. In this paper, we propose a latent fingerprint enhancement method based on DenseUNet. First, to generate the training data, the high-quality fingerprints are overlapped with the structured noises. Then, a deep DenseUNet is constructed to transform the low-quality fingerprint image into the high-quality fingerprint image by pixels-to-pixels and end- to-end training. Finally, the whole latent fingerprint is iteratively enhanced with the DenseUNet model to achieve the image quality requirement. Experiment results and comparison on NIST SD27 latent fingerprint database are presented to show the promising performance of the proposed algorithm.