AI-Assisted Forensic Evidence Recognition From Crime Scene Images: Evaluation of Deep Learning Object Detection Models

Abstract

Comprehensive evidence recognition is vital to crime scene investigation. Time pressure and scene complexity often cause missed evidence. Missed or misidentified evidence can permanently compromise a case. This study evaluates deep learning object detection for forensic evidence recognition. It forms phase one of C-SAHAI, a human-AI programme. C-SAHAI supports forensic practitioners during on-scene evidence analysis. A domain-specific dataset of 7,028 images was compiled. Images were drawn from simulated scenes and open repositories. Images span 18 forensic evidence classes after taxonomy correction. Classes include firearms, edged weapons, biological traces, and micro-evidence. Two architectures were trained: YOLOv8m and Faster R-CNN. Both models were evaluated at 35 and 100 epochs. Faster R-CNN showed the most balanced detection performance overall. At 100 epochs, its F1-score reached 0.705. Precision was 0.747, and recall was 0.668. YOLOv8m achieved higher mAP@50, reaching 0.453 at 100 epochs. However, its recall remained lower, at only 0.446. This trade-off reflects each architecture's underlying design philosophy. Extended training modestly improved both tested architectures overall. Neither model showed performance degradation across the tested range. Firearms and edged weapons were detected more reliably. Micro-evidence such as hair remained consistently difficult to detect. This reflects a known constraint of current detection architectures. These findings establish a baseline for AI-assisted evidence recognition. The approach supports, rather than replaces, expert forensic judgment.