面向无人机智能搜救场景的AI红外视觉超分辨重建方法研究

Research on an AI-Based Infrared Vision Super-Resolution Reconstruction Method for Intelligent UAV-Based Search and Rescue Scenarios

摘要: 在山地搜救、地震、洪涝等野外应急场景中，搭载红外成像设备的无人机可实现全天候大范围巡检作业，但飞行抖动、远距离成像衰减、水雾烟尘、植被遮挡及野外噪声等多重干扰，会导致红外图像清晰度下降、远距离人体热源特征弱化，极易引发搜救漏检问题。针对野外低分辨率红外图像细节丢失、无人机机载嵌入式算力有限、复杂工况泛化能力薄弱三大工程难题，本文开展专项算法优化研究。研究构建涵盖大气透射、水雾散射、植被遮挡的多等级复合退化模型，剖析传统滤波、通用轻量化网络、多模态融合重建三类技术方案的数学机理，并融合深度可分离卷积、多尺度热源注意力机制与像素–梯度联合边缘损失函数，设计出适配机载设备的轻量化红外超分辨网络。研究通过构建野外复合退化双数据集，依托PSNR、SSIM、推理耗时等多维指标，完成算法横向对比与模块消融实验，并结合沙盘仿真、户外无人机航拍及ARM嵌入式机载硬件开展实景实测。实验结果显示，该网络在基础退化场景下重建图像PSNR达36.10 dB、SSIM为0.9680，复合恶劣工况下PSNR仍可达34.07 dB，精度稳定性优于各类对比算法；机载设备单帧推理耗时86 ms，满足无人机端侧快速处理需求。

Abstract: In outdoor emergency scenarios such as mountain search and rescue operations, earthquake response, and flood mitigation, drones equipped with infrared imaging devices can perform large-scale inspection tasks under all weather conditions. However, multiple interference factors—including flight vibration, attenuation of long-distance imaging signals, water mist or dust, vegetation obscuration, and field noise—can lead to reduced clarity of infrared images and weakened detection of human thermal signatures at long distances, thereby significantly increasing the risk of missed detection in search and rescue operations. To address three major engineering challenges—loss of detail in low-resolution outdoor infrared images, limited onboard computational power of drones, and weak generalization capabilities under complex operating conditions—this paper conducts specialized research on algorithm optimization. The study constructs a multi-level composite degradation model that incorporates atmospheric transmission, water mist scattering, and vegetation obscuration; analyzes the mathematical mechanisms underlying three technical approaches-traditional filtering, lightweight universal networks, and multimodal fusion-based reconstruction; and integrates deep separable convolutions, a multi-scale thermal source attention mechanism, and a pixel-gradient joint edge loss function to design a lightweight infrared super-resolution network tailored for airborne platforms. Through the construction of two composite degradation datasets for outdoor environments, this study employs multi-dimensional evaluation metrics-including PSNR, SSIM, and inference time-to conduct comprehensive comparative analyses of the proposed algorithms and module ablation experiments; furthermore, field validation tests are performed using sandbox simulations, outdoor drone aerial photography, and ARM-based embedded airborne hardware. Experimental results demonstrate that the network achieves a PSNR of 36.10 dB and an SSIM of 0.9680 for image reconstruction under basic degradation scenarios; even under composite severe operating conditions, the PSNR remains at 34.07 dB, with superior accuracy stability compared to all benchmark algorithms; the onboard equipment requires only 86 ms for single-frame inference, meeting the quick processing requirements of UAVs.