基于机载的红外动态目标视频实时超分辨率重建  

Real-time super-resolution for infrared dynamic object video based on airborne platform

作　　者：朱德燕 徐家一 敖咏琪 ZHU Deyan;XU Jiayi;AO Yongqi(College of Astronautics,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China;Key Laboratory of Space Photoelectric Detection and Sensing of Industry and Information Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)

机构地区：[1]南京航空航天大学航天学院,江苏南京210016 [2]南京航空航天大学空间光电探测与感知工信部重点实验室,江苏南京210016

出　　处：《光学精密工程》2025年第5期818-828,共11页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.62105145);南京航空航天大学研究生科研与实践创新计划资助项目(No.xcxjh20221508,No.xcxjh20231505)。

摘　　要：为了提高机载红外成像系统对动态目标的远距离探测与识别能力,提出一种基于循环残差神经网络的红外视频超分辨率重建方法。该方法针对机载红外成像系统的实际退化过程,结合动态目标的运动信息,通过优化网络架构有效提升视频重建质量。首先,分析了包括下采样、运动模糊及噪声干扰在内的红外视频退化过程并基于此构建了低分辨率数据集,介绍了循环残差神经网络,该网络能够有效提取并传递动态目标的运动信息,从而恢复目标的形状、轮廓和细节纹理。采用跳跃级联残差结构改进模型主干,保证流畅信息流的同时使其更适合处理长视频序列,且有效避免了模型在训练过程中梯度消失。进一步,通过调整残差块的数量和各层卷积核的数量,优化了网络的表达能力和计算效率。此外,提出一种结合Charbonnier损失和高频信息损失(HFLoss)的损失函数共同监督,用于提升重建图像中高频细节的恢复效果。实验结果表明:所提出的重建方法在公开和实测红外数据集上均可实现动态目标的2倍超分辨率,PSNR值高于40 dB,SSIM值大于0.92,重建速率不低于45 frame/s。结合分辨率测试靶标与红外变焦成像系统准确标定了系统角分辨率,验证了重建方法在提升系统角分辨率方面的优势,系统角分辨率提升1.43倍。该方法能够满足机载成像系统高实时性和重建质量的要求。To enhance the long-range detection and recognition capabilities of airborne infrared imaging systems for dynamic targets,a video super-resolution reconstruction method based on a recurrent residual neural network is proposed.This method addresses the degradation process inherent to airborne infrared imaging systems and incorporates motion information from dynamic targets to improve video reconstruction quality through optimization of the network architecture.Initially,the degradation process of infrared video is analyzed,encompassing downsampling,motion blur,and noise interference,leading to the construction of a low-resolution dataset reflective of these factors.Subsequently,the recurrent residual neural network is introduced,which effectively extracts and propagates motion information of dynamic targets,thereby restoring the shape,contours,and intricate details of the targets.A skip-connected residual structure is implemented to enhance the network backbone,ensuring smooth information flow while increasing suitability for processing extended video sequences and effectively mitigating the gradient vanishing problem during training.Furthermore,by adjusting the number of residual blocks and the convolution kernel sizes within each layer,the expressive power and computational efficiency of the network are optimized.Additionally,a novel loss function is proposed,which combines Charbonnier loss and high-frequency information loss(HFLoss)for joint supervision,facilitating improved recovery of high-frequency details in the reconstructed images.Experimental results demonstrate that the proposed method achieves 2 times superresolution for dynamic targets on various publicly available and experimentally collected infrared datasets,yielding a PSNR exceeding 40 dB and an SSIM above 0.92,with a reconstruction rate of no less than 45 frame/s.Moreover,the system's angular resolution is accurately calibrated utilizing a resolution test target alongside an infrared zoom imaging system,substantiating the advantages of the propos

关 键 词：计算机视觉 视频超分辨率 深度学习 循环神经网络 深度残差网络 红外动态目标 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]