Face comparison in forensics A deep dive into deep learning and likelihood rations

This thesis explores the transformative potential of deep learning techniques in the field of forensic face recognition. It aims to address the pivotal question of how deep learning can advance this traditionally manual field, focusing on three key areas: forensic face comparison, face image quality assessment, and likelihood ratio estimation. Using a comparative analysis of open-source automated systems and forensic experts, the study finds that automated systems excel in identifying non-matches in low-quality images, but lag behind experts in high-quality settings. The thesis also investigates the role of calibration methods in estimating likelihood ratios, revealing that quality score-based and feature-based calibrations are more effective than naive methods. To enhance face image quality assessment, a multi-task explainable quality network is proposed that not only gauges image quality, but also identifies contributing factors. Additionally, a novel images-to-video recognition method is introduced to improve the estimation of likelihood ratios in surveillance settings. The study employs multiple datasets and software systems for its evaluations, aiming for a comprehensive analysis that can serve as a cornerstone for future research in forensic face recognition.