Detection and Identification of Unexploded Ordnance Using a Two-Step Deep Learning Methodology

Localisation and disposal of unexploded ordnance (UXO) is a crucial task that can save the lives of both military personnel and civilians. In comparison to immediate post-war intervention situations, EOD (Explosive Ordnance Disposal) teams can now leverage emerging technologies based on computer vision architectures, mitigating the perceived risks associated with hands-on inspection of ammunition. The paper analyzes the use of new convolutional neural network architectures, in detection and identification of unexploded ordnance by combining specialised domain knowledge with computer vision models and methods. Additionally, it presents image preprocessing methods, research techniques, results, and conclusions. Moreover, we propose a complementary approach to previous research, often based on the interpretation of external sensor signals, representing the missing link in a comprehensive and extensive identification. Standardized metrics such as mean average precision, precision, recall, and F1-score are reported to evaluate the outcomes. Results, using the YOLOv8 architecture, achieve up to 80.8% mAP for the binary classification task (detection problem) and up to 90.6 % mAP performance for the subsequent identification task. The work aims to offer a baseline study and new perspective on addressing a highly significant issue: mitigating by computer vision the risks associated with unexploded ordnance.