Deep Learning-Based Camouflage Detection for Anti-Personnel Mine Identification in Natural Environments

Abstract

The detection of anti-personnel mines in natural environments remains a critical humanitarian and technological challenge due to the high visual similarity between explosive devices and their surrounding backgrounds. This study presents a deep learning-based camouflage detection framework for the identification and segmentation of PFM-1 anti-personnel mine surrogates embedded in visually homogeneous outdoor scenes. The proposed approach employs a Deep Camouflage Detection Network (DCDN) designed to extract multi-scale contextual features and enhance boundary sensitivity under low-contrast conditions. A dedicated dataset was constructed using 3D-printed PFM-1 surrogates positioned in vegetated environments under varying illumination conditions, viewing angles, occlusion levels, and object scales. The network architecture integrates a pretrained convolutional backbone, multi-scale feature aggregation modules, and a  composite loss function (consisting of Binary Cross-Entropy and Dice loss) to address class imbalance and weak edge contrast.
Experimental evaluation on an independent test set demonstrates robust segmentation performance, achieving a mean Intersection over Union (IoU) of 91.8% and a Dice coefficient of 95.7%. Precision and recall values of 96.3% and 94.9%, respectively, confirm a balanced detection capability with limited false positives. Stratified analysis indicates stable performance under illumination variability and partial occlusion, while ablation studies highlight the importance of multi-scale aggregation and region-aware loss optimization. The results confirm that deep camouflage-aware segmentation architectures provide reliable detection of low-contrast objects in complex natural environments.