A Fine-Grained Detection Network Model for Soldier Targets Adopting Attack Action

Owing to its ability to provide more accurate and detailed battlefield situational information, fine-grained detection research on soldier targets is of significant importance for military decision-making and firepower threat assessment. To address the issues of low detection accuracy and inaccurate classification in the fine-grained detection of soldier targets, we propose a fine-gain soldier target detection model based on the improved YOLOv8 (You Only Look Once v8). First, we developed a multi-branch feature fusion module to effectively fuse multi-scale feature information and used a dynamic deformable attention mechanism to help the detection model focus on key areas in deep-level features. Second, we proposed a decoupled lightweight dynamic head to extract the position and category information of soldier targets separately, effectively solving the problem of misclassification of soldier targets' attack actions under different poses. Finally, we used the Inner Minimum Points Distance Intersection over Union (Inner-MPDIoU) to further improve the convergence speed and accuracy of the network model. The proposed improvements are evaluated through comparative experiments conducted in published twenty-six test groups, and the effectiveness of the proposed method is demonstrated. Compared with the original model, our method achieved a detection precision of 78.9%, a 6.91% improvement; the mAP@50 (mean Average Precision at 50) was 79.6%, a 3.51% increase; and an mAP@50-95 of 63.8%, a gain of 5.28%. The proposed method achieves high precision and recall while reducing the computational complexity of the model, thereby enhancing its efficiency and robustness for fine-grained soldier target detection.