International Journal of Scientific Research and Engineering Development

Face detection is a critical task in digital forensic investigations, supporting essential activities like suspect tracking, victim identification, and multimedia triage. Traditional and even modern deep learning methods often falter under forensic conditions, due to factors like low resolution, motion blur, occlusion, and suboptimal lighting. This study investigates the efficacy of optimized YOLO-based architectures (YOLOv8, YOLOv10, YOLOv12) for forensic face detection by retraining nano, small, and medium variants of each model on a refined subset of the WIDER FACE dataset. The proposed preprocessing approach eliminates images below 640×640 pixels to enhance learning efficiency and detection accuracy. Experimental results demonstrate significant improvements across precision, recall, and inference speed metrics compared to pretrained baselines, with YOLOv12 achieving superior latency and precision scores. The findings highlight the importance of task-specific training and dataset refinement in digital forensics.