A model for person re-identification using deep learning techniques: a case of Ugandan school candidates

Abstract

Person Re-Identification (Re-ID) is a critical domain in computer vision, with signifi- cant applications in security and public safety. This thesis addresses the unique and com- plex challenge of Re-ID within the educational sector, specifically focusing on the Uganda National Examinations Board (UNEB) dataset. The core problem is the need for reli- able re-identification of school candidates across a minimum four-year age gap (Primary Leaving Examination - PLE to Uganda Certificate of Education - UCE), which causes se- vere Age-Variant Feature Drift. Existing general-purpose Re-ID models are insufficiently robust to these extreme age-related changes, leading to a high risk of misidentification, spoofing, and impersonation during exam registration. To solve this problem, the research study proposes the Automated Person Re-Identification (APRe-ID) model, a novel deep learning architecture designed to generate age-invariant feature embeddings from facial images. The study utilizes a proprietary UNEB dataset of 299,744 data points of student facial features, captured four years apart, to rigorously test the proposed solution. The methodology involves comprehensive data pre-processing with image augmentation, robust face detection using MTCNN, and a deep learning model incorporating the strengths of Deep Metric Learning (DML) and Attention-Based Ar- chitectures. The specific objectives were to investigate optimal facial feature extractors, design the APRe-ID model, and evaluate its performance against the state-of-the-art. Our empirical results show that APRE-ID achieves Precision of 99%, Recall of 99% and F1-Score of 99% as an ensemble of Random forest, Extreme Gradient Boosting, and Sup- port Vector Machine.