ABSTRACT
Latent fingerprints are the fingerprints which are left unintentionally on a surface while touching it. These are of great interest for the forensics experts for criminal identification. Latent fingerprints usually possess high nonlinear distortion. Furthermore, these may be overlapping with background text or other fingerprints. These fingerprints can be extracted from different surfaces leading to the varying background. The presence of structured and unstructured background noise adversely affects minutiae (ridge bifurcation/ridge ending) extraction in latent fingerprints which in turn leads to poor matching performance. A latent fingerprint enhancement algorithm removes the background noise and predicts the missing ridge information. It also improves the ridge clarity which helps to improve minutiae extraction and thereby improving matching performance. Traditionally, latent fingerprints are enhanced by approximating the orientation field and then applying contextual filtering using the approximated orientations. However, recently the attention has been shifted towards developing models which can directly denoise the fingerprints and reconstruct the missing ridge structure without explicitly estimating the orientation field.
Inspired by the success of Generative Adversarial Network (GAN) in image processing applications, we propose a GAN-based latent fingerprint enhancement model. However, one of the key issues with GANs is that they are difficult to train. Through this work, we contribute our efforts towards sharing details on successfully training a GAN. The proposed latent fingerprint enhancement model preserves the ridge structure including minutiae. We discuss the role of training data, i.e., various noise models which should be considered for modelling a latent fingerprint, during training a GAN. In addition to this, we discuss the significance of choice of loss function and the role of hyper-parameters such as batch size, weight of each loss term, and number of epochs for training the GAN. We evaluate the proposed enhancement model on publicly available latent databases: Indraprastha Institute of Information Technology Delhi Multi-sensor Optical and Latent Fingerprint (IIITD-MOLF) and Indraprastha Institute of Information Technology Delhi Multi-surface Latent Fingerprint (IIITD-MSLF).
