Fingermark quality assessment framework with classic and deep learning ensemble models

摘要

The quality assessment of fingermarks (latent fingerprints) is an essential part of a forensic investigation. It indicates how valuable the fingermarks are as forensic evidence, it determines how they should be further processed, and it correlates with the likelihood of successful identification, i.e., finding a matching fingerprint in a reference database. Since the environments in which fingermarks are found are not controlled, this task proves challenging even with modern machine learning solutions. In this work, we propose a predictive framework for automated fingermark quality assessment (AFQA). With this iteration of AFQA, we bridge the gap between the classic machine learning approach with handcrafted features and the modern deep learning paradigm, evaluate the advantages and disadvantages of these methodologies, and provide the rationale and direction for future development of AFQA methods. We present a significantly improved AFQA toolbox and provide a quality aggregation method capable of fusing together multiple predicted quality values from an ensemble of quality assessment models. The proposed ensemble approach provides improved prediction performance while reducing processing time compared to existing state-of-the-art solutions.