Improving Identification by Removing Obstructive Minutiae From Latent Fingerprints

Latent fingerprint identification is essential in forensic science for linking suspects to crime scenes or, conversely, confirming a person’s innocence. However, latent fingerprints often are partial prints with undesirable characteristics such as noise or distortion. Due to these characteristics, identifying the physical details of a latent fingerprint, known as minutiae, is a complex task. Recent publications found that there are minutiae in latent fingerprints that, when removed, increase the matching score. We have defined this type of minutia as obstructive. The importance of obstructive minutiae lies in their ability to increase the identification rate when they are identified and removed. In this work, we propose a new set of features to describe obstructive minutiae in latent fingerprints. Using this set of features, we have built datasets that describe latent fingerprints from which a subset of minutiae has been removed. Additionally, we have evaluated a set of multi-class classifiers trained with our datasets to predict if there are obstructive minutiae in a latent fingerprint. Finally, we developed two novel algorithms to identify and remove, in latent fingerprints, the obstructive minutia that generates the maximum increase in the matching score according to our set of classifiers. We employed Cumulative Match Characteristic (CMC) curves to compare the relative change of identifying an initial latent fingerprint versus a latent fingerprint with the removed obstructive minutia with the maximum increase in the matching score. We achieved an identification rate of 88.2% at Rank 1 using the XGBoost classifier on the NIST SD27 database. Similarly, the Random Forest classifier achieved an identification rate of 73.5% at Rank 1 on the NIST SD302 database.