Optimizing military target recognition in urban battlefields: an intelligent framework based on graph neural networks and YOLO

In urban battlefield environments, the rapid movement and frequent occlusion of military targets often result in lower detection accuracy. To address this issue, this study proposes an innovative approach that combines graph neural networks with the YOLO model to overcome the slow detection speed and low accuracy due to blurriness in existing models. We first detect the targets, then enhance model performance by introducing intelligent reasoning and optimization processes at the output stage, allowing the model to reassess object confidence based on spatial relationships between objects. A graph relationship model is constructed from the detection results and input into the adjusted SeHGNN network. The SeHGNN network learns complex relationships between targets and recalculates confidence scores. Experimental results show significant improvements in mAP@0.50, demonstrating the effectiveness of this method. By integrating traditional object detection techniques with the knowledge reasoning capabilities of graph neural networks, this approach substantially enhances the model’s performance in detecting military targets in urban battlefield scenarios.