基于多尺度特征融合网络的机载雷达前视成像方法

Airborne Radar Forward-Looking Imaging Method Based on Multi-Scale Feature Fusion Network

摘要:

由于机载雷达前视区域的多普勒梯度趋于零，传统基于多普勒效应的合成孔径雷达（synthetic aperture radar，SAR）等高分辨率成像方法在该区域方位向分辨率严重恶化，形成所谓的“成像盲区”。针对这一瓶颈问题，提出了一种基于仿真数据集的机载雷达前视成像新方法，并构建了多尺度特征融合网络（multi-scale feature fusion network，MFF-Net）。该网络采用U形编码器-解码器架构，创新性地设计了多尺度输入与输出机制，能够从雷达脉冲压缩后的原始回波中端到端地重建高分辨率前视场景图像。针对真实场景下高分辨率前视成像标签数据稀缺的问题，通过对高分辨率SAR图像进行回波仿真，构建了大规模、高可靠的训练数据集。实验结果表明，无论在点目标还是场景目标成像中，MFF-Net均表现出卓越的超分辨成像性能，其成像结果在清晰度、目标结构保真度方面均显著优于传统实孔径成像和单脉冲成像方法，验证了该方法在前视成像领域的有效性和应用潜力。

Abstract:

In the airborne radar forward-looking region, the Doppler gradient approaches zero, causing the azimuth resolution of conventional high-resolution imaging methods like synthetic aperture radar (SAR), which rely on the Doppler principle, to degrade severely. This results in a so-called ‘blind zone’. To address this bottleneck, a novel method for airborne radar forward-looking imaging based on a simulated dataset was proposed, utilizing a multi-scale feature fusion network (MFF-Net). The network employed a U-shaped encoder-decoder architecture and featured an innovative multi-scale input and output mechanism. The high-resolution forward-looking scene images were directly reconstructed from the pulse-compressed raw echo in an end-to-end fashion. Given the scarcity of high-resolution labelled forward-looking data in real-world scenarios, a large-scale and reliable training dataset for the network was generated by performing an echo simulation on high-resolution SAR images. Experimental results demonstrate that MFF-Net delivers excellent super-resolution imaging performance for both point and scene targets. The imaging results show significant improvements in terms of clarity and target structure fidelity compared to those of the conventional real-aperture and monopulse imaging methods, validating the effectiveness and application potential of the proposed method in the field of forward-looking imaging.