基于改进YOLOv11n的无人机红外目标检测算法

Infrared Target Detection Algorithm in UAV Imagery Based on Improved YOLOv11n

摘要: 面向无人机红外图像中目标尺度小、对比度低与边界模糊等问题，本文提出了一种基于YOLOv11n模型的多尺度注意力机制优化方法。首先，在引入小目标检测层的基础上，融合多分支与双向金字塔思想构建双向多分支辅助特征金字塔网络，通过可学习权重自适应融合各层特征，增强微小目标表征。其次，在检测头侧采用动态注意力检测头，从尺度、空间与通道三方面进行协同建模，提升关键区域聚焦与特征利用效率。最后，提出NWD-Inner-MPDIoU组合损失函数，协同提升低重叠、边界不清条件下的定位稳定性。在HIT-UAV红外小目标数据集上进行系统实验评估，结果表明：所提方法mAP50达92.8%，相比基线模型提升2.2%，且召回率与准确率分别提高1.6%和0.6%。同时，模型仅小幅增加复杂度，整体仍保持轻量化与可部署性。综上，本文方法在保证效率的同时有效提升了无人机红外目标的检测质量，为后续扩展研究提供了可靠的技术基础。

Abstract: To address the challenges of small target scale, low contrast, and blurred boundaries in UAV infrared imagery, this paper proposes a YOLOv11n-based method enhanced with multi-scale attention. First, on top of an additional small-object detection layer, we integrate multi-branch and bidirectional pyramid designs to construct a bidirectional multi-branch auxiliary feature pyramid network, in which learnable weights adaptively fuse multi-level features to strengthen the representation of diminutive targets. Second, we employ a dynamic attention-based detection head that jointly models scale, spatial, and channel dimensions to improve focus on salient regions and the efficiency of feature utilization. Finally, we introduce a compound loss, NWD-Inner-MPDIoU, which jointly enhances localization stability under low-overlap and ambiguous-boundary conditions. Comprehensive experiments on the HIT-UAV infrared small-target dataset show that the proposed method achieves an mAP50 of 92.8%, outperforming the baseline by 2.2%; recall and precision are improved by 1.6% and 0.6%, respectively. Meanwhile, the model incurs only a slight increase in complexity and remains lightweight and deployment-ready. Overall, the proposed approach effectively improves infrared UAV target detection while maintaining efficiency, providing a reliable technical foundation for subsequent research.