摘要
Accurate object detection is essential for military surveillance, yet standard evolutionary hyperparameter optimization (EHPO) for models such as You Only Look Once (YOLO) often suffers from computational inefficiency and evaluation bias, primarily due to its memoryless design and reliance on single data splits. To address these limitations, this study proposes Greedy and Dynamic Evolutionary Hyperparameter Optimization (GD-EHPO), a general-purpose framework that integrates Greedy K-fold Evaluation (GKE) for robust fitness estimation and Dynamic Weight Inheritance (DWI) to enable continuous knowledge transfer across generations. Validated on the Military Aircraft Detection Dataset using YOLOv11s, GD-EHPO consistently outperforms conventional EHPO frameworks across multiple evolutionary algorithms. In particular, when combined with Simplified Swarm Optimization (SSO), GD-EHPO achieves a 4.81% improvement in mAP@0.5 over the default baseline and a 2.59% gain over standard EHPO, while demonstrating faster convergence and stronger cross-version generalization. These findings indicate that GD-EHPO provides an effective and scalable solution for optimizing real-time object detection models under resource-constrained deployment conditions. © 2026 Elsevier B.V.