FinSNet: End-to-End Separation of Overlapped Fingerprints Using Deep Learning

Latent fingerprints found at a crime scene are an important evidence in a criminal investigation. However, latent fingerprints are often damaged by the surface of an object or obscured by other latent fingerprints. Herein, we propose an end-to-end overlapped fingerprint separation method using a deep learning algorithm. The neural network, FinSNet, takes an overlapped fingerprint image as an input and is trained to restore the component fingerprint from the input directly. Our network can remove redundant fingerprints and background images with only a single inference, resulting in improved computational efficiency compared to conventional methods that require background removal to be preceded separately. First, we introduce strategies to design the network and build a training dataset. Next, we demonstrate results of separating real-world fingerprints, using Tsinghua overlapped fingerprint datasets. We then evaluate our proposed method using commercial fingerprint identification software to perform fingerprint matching. Finally, we thoroughly analyze our method in terms of the training dataset, background removal, and angular deviation of overlapped fingerprints.