Performing recognition tasks using distorted fingerprint samples is often challenging for automated identification systems due to poor quality, distortion, and partially missing information from the input samples. Photometric and geometric distortions are the major factors that degrade the quality of fingerprints. Photometric distortion causes missing areas and spurious minutiae points while geometric distortion manipulates the spatial location of features. In this proposal, the researcher aims to incorporate deep learning models to rectify both the photometric and geometric distortion of fingerprints in a unified framework. The researcher proposes a four stage prject. First, the researcher proposes to exploit the prior ridge information of clean fingerprints using principal component analysis to develop a statistical model for the geometric distortion of distorted fingerprints. Second, the resarcher will use the statistical model of distortion to train a deep model capable of estimating the distortion parameters of the input samples and rectify the distortion. Third, the researcher will develop a conditional generative adversarial network (cGAN) to rectify the photometric distortion using a data-driven methodology that estimates missing areas based on ridge information of clean fingerprints. In addition, the cGAN model harnesses the information about the geometric distortion estimated in the previous stage to reduce the possibility of generating spurious minutiae. Fourth, the researcher will develop a probabilistic model which can predict the probability of two rectified latent samples belonging to a same ID. To achieve this goal,the applicant will train this model using synthetic partial latent fingerprint samples by deliberately distorting a large number of clean fingerprints. Combining all four models together, the researcher seeks to develop a unified fingerprint distortion rectification model capable of rectifying both the geometric and photometric distortion and estimating the correlation of latent samples collected from a crime scene. The researcher will evaluate the proposed method in combination with two different fingerprint matching algorithms on several publicly available latent fingerprint datasets.. Matching performance of the reconstructed latent fingerprints to rolled fingerprints will be evaluated using two matching techniques, the VeriFinger 7.0 SOK  and the NIST Biometrics Image Software (NBIS). The cumulative match curves (CMCs) and latent fingerprint image quality (LFIQ) criteria will be used to assess the value of our proposed cGAN fingerprint restoration methods.
Note: This project contains a research and/or development component, as defined in applicable law," and complies with Part 200 Uniform Requirements - 2 CFR 200.210(a)(14). CA/NCF