跳转至

使用合成数据训练热红外图像中的无人机检测

文章背景与核心概要

在地对空（G2A）中长波红外（MWIR/LWIR）无人机检测任务中，研究长期面临着纹理信息缺失、传感器噪声、热对比度微弱以及带标注真实世界数据严重匮乏等挑战。本文探讨了一种“合成数据优先”的训练策略，即将合成场景生成与真实数据的微调有机结合，为解决红外视觉领域的样本难题提供了一种极具前景的方案。

研究的核心结论表明：合成数据能够为学习初始对象特征提供极为有效的坚实基础；同时，即使是少量的真实红外图像对于实现可靠部署也至关重要，它能显著缩小域间差距。此外，实验证明数据集的对齐程度对检测性能的影响远大于模型规模的扩展。特征空间的语义对齐是模型性能最强的预测指标，而熵和动态范围等辐射度特性则能进一步增强检测的鲁棒性。

摘要

Ground-to-Air (G2A) drone detection in medium- and long-wave infrared (MWIR/LWIR) imagery is challenging due to reduced texture information, sensor noise, weak thermal contrast, and the scarcity of annotated data. This work investigates a synthetic-first training strategy that combines synthetic scene generation with fine-tuning on real data. We show that synthetic data provides an effective basis for learning initial object representations, while real in-domain thermal imagery is still essential for reliable deployment. Even small amounts of real IR data substantially reduce domain gaps. Our experiments indicate that dataset alignment has a stronger impact on performance than model scale. Finally, our analysis of the dataset suggests that semantic alignment in feature space is the strongest predictor of model performance, while radiometric properties such as entropy and dynamic range also contribute to detection robustness. This work provides a foundation for combining synthetic and real IR data for effective G2A drone detection.

在地对空（G2A）中长波红外（MWIR/LWIR）图像中的无人机检测面临诸多挑战，包括纹理信息减少、传感器噪声、热对比度微弱以及缺乏带标注数据。本研究探讨了一种“合成数据优先”的训练策略，该策略将合成场景生成与真实数据的微调相结合。我们表明，合成数据为学习初始目标表征提供了有效基础，而真实的域内热红外图像对于可靠部署仍然至关重要。即使少量的真实红外数据也能显着减少域差距。我们的实验表明，数据集对齐对性能的影响比模型规模更大。最后，我们对数据集的分析表明，特征空间中的语义对齐是模型性能的最强预测指标，而熵和动态 range 等辐射特性也有助于提高检测的鲁棒性。这项工作为结合合成和真实红外数据进行有效的 G2A 无人机检测奠定了基础。

总结

Ground-to-Air (G2A) drone detection using medium- and long-wave infrared (MWIR/LWIR) imagery is heavily hindered by reduced textures, sensor noise, weak thermal contrast, and a lack of annotated real-world data. This paper explores a synthetic-first training strategy—combining synthetic scene generation with fine-tuning on real data.

Key takeaways from the research include: * Initial Representations: Synthetic data provides a highly effective foundation for learning initial object features. * Real Data is Essential: Even small quantities of real infrared (IR) imagery are critical for reliable deployment and drastically reduce domain gaps. * Alignment over Scale: Dataset alignment has a significantly greater impact on detection performance than model scaling. * Performance Predictors: Semantic alignment in feature space serves as the strongest predictor of model performance, while radiometric properties (such as entropy and dynamic range) further enhance detection robustness.

使用中长波红外（MWIR/LWIR）图像的地对空（G2A）无人机检测严重受到纹理减少、传感器噪声、微弱热对比度以及缺乏带标注真实世界数据的阻碍。本文探讨了一种合成数据优先的训练策略——将合成场景生成与真实数据微调相结合。

该研究的主要结论包括： * 初始表征： 合金数据为学习初始目标特征提供了高效的基础。 * 真实数据至关重要： 即使是少量的真实红外（IR）图像对于可靠部署也至关重要，并且能大幅缩小域差距。 * 对齐优于规模： 数据集对齐对检测性能的影响显著大于模型规模的扩大。 * 性能预测指标： 特征空间中的语义对齐是模型性能的最强预测指标，而辐射特性（如熵和动态范围）则进一步增强了检测的鲁棒性。

文档详情

  • arXiv Identifier: arXiv:2608.17799 [cs.CV]
  • Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
  • Secondary Subjects: Artificial Intelligence (cs.AI), Emerging Technologies (cs.ET), Robotics (cs.RO)
  • Submission Date: August 18, 2026
  • Conference Context: To be presented at SPIE: Sensors + Imaging, Artificial Intelligence for Security and Defence Applications IV
  • License: Creative Commons Attribution 4.0 International license icon
  • arXiv 标识符： arXiv:2608.17799 [cs.CV]
  • 主要主题： 计算机视觉与模式识别 (cs.CV)
  • 次要主题： 人工智能 (cs.AI)、新兴技术 (cs.ET)、机器人学 (cs.RO)
  • 提交日期： 2026年8月18日
  • 会议背景： 将在 SPIE: Sensors + Imaging, Artificial Intelligence for Security and Defence Applications IV 上发表
  • 许可协议： 知识共享署名 4.0 国际许可协议 license icon

作者

  • Tanel Liiv
  • Sander Soodla
  • Nzamba Bignoumba
  • Alma M. Liezenga
  • Toomas Pruuden
  • Tanel Liiv
  • Sander Soodla
  • Nzamba Bignoumba
  • Alma M. Liezenga
  • Toomas Pruuden

资源与链接