Wang Haixia, Wang Hailong, Han Fen
Department of Mechanical and Electrical Engineering, Hetao College, Bayannur, 015000, China.
Dengkou County Sand Prevention and Control Bureau, Bayannur, 015200, China.
Sci Rep. 2025 Jul 2;15(1):23344. doi: 10.1038/s41598-025-06070-1.
The objective of this paper is to address the issue of the inadequate detection accuracy of UAVs operating at low-altitudes in conditions of weak thermal signals. To this end, an enhanced small target detection model has been proposed, which integrates density-peak clustering and an artificial bee colony optimization mechanism. In the study, multi-stage preprocessing is first performed on the infrared images. By combining Butterworth low-pass filtering, local background subtraction, and exponential high-pass filtering, weak target templates are extracted. Then, the DBABC algorithm is utilized to achieve efficient clustering and accurate identification of high-density areas. To adapt to the dynamic characteristics of the target, size change perception and environmental perturbation correction mechanisms have been introduced. The experimental results showed that the model achieved the highest detection accuracy of 91.66% and 90.38% on the FLIR and KAIST datasets, respectively, with a recall rate of over 89.6%. The model maintained a signal-to-noise ratio gain of over 23.19 dB in long-distance detection from 150 to 200 m, and the detection delay was controlled below 0.5 dB at resolutions of 3840 × 2160 and 7680 × 4320. The detection delay was controlled within 28-36 ms, which was significantly better than the advanced model. Further tests in complex weather and micro-temperature difference environments verified the robustness and wide adaptability of the proposed method under rain and fog interference, high background noise, and very low thermal contrast conditions. This study enriches the application framework of density-driven intelligent optimization strategies in the field of infrared small target detection. Moreover, it provides theoretical support and technical path for low-altitude unmanned aerial vehicles to achieve efficient and stable infrared target detection in tasks such as security patrols, disaster search and rescue, and border monitoring.
本文的目的是解决无人机在弱热信号条件下低空运行时检测精度不足的问题。为此，提出了一种增强型小目标检测模型，该模型集成了密度峰值聚类和人工蜂群优化机制。在研究中，首先对红外图像进行多阶段预处理。通过结合巴特沃斯低通滤波、局部背景减法和指数高通滤波，提取弱目标模板。然后，利用DBABC算法实现对高密度区域的高效聚类和准确识别。为了适应目标的动态特性，引入了尺寸变化感知和环境扰动校正机制。实验结果表明，该模型在FLIR和KAIST数据集上分别实现了91.66%和90.38%的最高检测精度，召回率超过89.6%。该模型在150至200米的远距离检测中保持了超过23.19 dB的信噪比增益，在3840×2160和7680×4320分辨率下检测延迟控制在0.5 dB以下。检测延迟控制在28 - 36毫秒内，明显优于先进模型。在复杂天气和微温差环境下的进一步测试验证了该方法在雨雾干扰、高背景噪声和极低热对比度条件下的鲁棒性和广泛适应性。本研究丰富了密度驱动智能优化策略在红外小目标检测领域的应用框架。此外，它为低空无人机在安全巡逻、灾害搜索救援和边境监测等任务中实现高效稳定的红外目标检测提供了理论支持和技术路径。
