抗干扰注意力赋能的YOLOv8在复杂背景遥感影像中的目标检测

Target Detection in Remote Sensing Images with Complex Background Based on Anti-Interference Attention Enhanced YOLOv8

摘要: 针对复杂场景下遥感目标检测存在的小目标易漏检、特征表征能力弱、检测精度与鲁棒性不足等问题，研究提出了一种改进的YOLOv8遥感目标检测模型。通过构建PCM像素关系模块增强局部上下文关联，提升小目标与弱特征目标的表达能力；设计MultiSEAM多尺度注意力机制，强化多尺度特征融合并抑制背景噪声；引入InnerShape‑IoU损失函数优化边界框回归，提高模型收敛性与定位精度。在UCAS‑AOD数据集上的实验表明，所提模型mAP@0.5达96.2%，与基准模型基本持平，参数量与推理速度均衡，在小目标、密集目标及复杂背景下鲁棒性更强。该模型在轻量化前提下实现性能提升，可为遥感监测、灾害评估、国土资源调查等领域提供有效技术支撑。

Abstract: To address the problems of small object missed detection, insufficient feature representation, and low detection accuracy in remote sensing object detection under complex scenarios, this study proposes an improved YOLOv8 model for remote sensing object detection. First, the PCM pixel context module is constructed to enhance local context association and improve feature expression for small and weak objects. Second, the MultiSEAM multi-scale attention mechanism is designed to strengthen multi-scale feature fusion and key information extraction while suppressing background noise. Meanwhile, the InnerShape-IoU loss function is adopted to optimize bounding box regression, enhancing model convergence and localization accuracy. Experimental results on the UCAS-AOD dataset show that the proposed model achieves 96.2% mAP@0.5, which is almost consistent with the baseline model, and maintains a good balance between parameters and inference speed. It exhibits stronger robustness for small objects, dense objects, and objects under complex backgrounds. The improved model achieves significant performance improvement while remaining lightweight, providing effective technical support for remote sensing monitoring, disaster assessment, land resource survey, and other fields.