U-Net based Deep Learning Architectures for Latent Fingerprint Segmentation

In this paper we propose to train different CNN-based encoder-decoder models for latent fingerprint segmentation. Model selection is conducted through hyper-parameters tweaking, and per-formance assessment, of individual models. Qualitative and quantitative evaluations, involving standard metrics, learning curves, and mask prediction versus ground truth, are reported using the NIST SD27 and NIST SD301 databases. The best deep learning model producing the lowest generalisation error on the segmentation task is highlighted.