Studies on Efficient Techniques for Enhancing Latent Fingerprint Recognition Systems

Abstract

Latent fingerprints are impressions that have inadvertently been left on materials at the scene of crime. In the forensic science literature and legal system, there have been growing investigation, analysis, and debate on the reliability of latent fingerprint recognition by latent fingerprint forensic experts. Incorrect conviction of innocent individuals due to mistakes in latent fingerprint matching is critical. Enforcement department uses latent fingerprint comparison increasingly to identify crimes and prosecute convicts. Latent investigators currently label the Region Of Interest (ROI) manually in latent fingerprints and utilize characteristic features manually detected from the ROI to search on fingerprint libraries to discover a small number of potential matches for further human examination. It is highly desirable to carry out latent fingerprint analysis in a fully automated manner due to the large number of law enforcement databases with plain and rolled fingerprints. For automated fingerprint image quality evaluation, quality enhancement, segmentation and matching, this thesis introduces deep learning (DL) models built in the framework of machine learning (ML). newlineAdditionally, methods for accelerating deep neural network convergence and improving the estimation of the correlation among a latent fingerprint photograph patches and their target class are put forth. This research aims to provide an end-to-end automated process that addresses the issues associated with latent fingerprint quality enhancement, segmentation, quality evaluation and matching using DL techniques. newlinevi newlineIn the first method, the proposed approach combines a sparse representation with multi-scale patching (ASR-MSP) and a total variation model. The animation and pattern components of the picture are separated into two categories by the TV model. The texture elements are defined as the information structure of minuscule patterns, while the cartoon elements are excluded as non-fingerprint characteristic patterns.