Latent Fingerprint Image Enhancement Based on Progressive Generative Adversarial Network
Xijie Huang, Peng Qian, Manhua Liu
Abstract
Latent fingerprints, a kind of fingerprints which are captured from the finger skin impressions at the crime scene, have been adopted to identify suspected criminals for a long time. However, poor latent fingerprint image quality owing to unstructured overlapping patterns, unclear ridge structure, and various background noise has brought a challenge to the recognition of latent fingerprints. Therefore, image enhancement is a crucial step for more accurate fingerprint recognition. In this paper, a latent fingerprint enhancement method based on the progressive generative adversarial network (GAN) is proposed. The powerful GAN structure provides an efficient translation from latent fingerprint to high-quality fingerprint. Our method consists of two stages: Progressive Offline Training (POT) and Iterative Online Testing (IOT). Progressive training makes our model not only focus on the local features such as minutiae but also preserve structure feature such as the orientation field. We extensively evaluate our model on NIST SD27 latent fingerprint dataset. With the help of orientation estimation task and progressive training scheme, our model achieves better recognition accuracy.