关键词: 军事飞机目标检测, YOLOv8, 深度特征提取, 动态上采样, 统一参数化

Abstract: Aiming at the problems of high leakage rate and insufficient feature interaction due to low contrast, small size and variable morphology in military aircraft target detection in the air-sky context, this study proposes an improved SDU-YOLOv8 network based on YOLOv8. The synergistic optimization of feature extraction, fusion and detection head is achieved by constructing SSGBlock deep feature extraction module, dynamically learnable Dy-RepGFPN feature fusion network, and parameter-sharing UCDN-Head detection head. Through experiments on a self-constructed military aircraft dataset, the SDU-YOLOv8 network outperforms the benchmark YOLOv8 mAP@0.5 Improved by 2.5%, 6.7% reduction in the amount of parameters, 9.9% reduction in the amount of computation, and significant enhancement in the detection robustness of small-size, low-contrast and deformation targets. The experimental results show that the proposed method achieves a balanced optimization of detection accuracy and efficiency while maintaining lightweight, providing an efficient solution for military aircraft detection in air and space reconnaissance scenarios.

Key words: military aircraft target detection ;YOLOv8, deep feature extraction, dynamic upsampling, unified parameterization