舰船电子对抗

针对数字射频存储器(DRFM)密集假目标信号在时频域难以精细定位的问题，围绕如何为后续DRFM分量抑制提供可解释的像素级空间先验展开研究。基于统一仿真平台，对目标、海杂波与DRFM密集假目标信号进行联合建模，将完整时域信号经短时傅里叶变换(STFT)得到的复值时频图作为主要分析对象，在同一时频平面上刻画目标、背景与DRFM信号的时频结构及能量分布特征。在此基础上，利用仿真中可获得的分量分解信息构造软掩码(SoftMask)、硬掩码(HardMask)、核心掩码(CoreMask)与有效掩码(ValidMask)4类掩码监督标签，以刻画DRFM分量的分级区域结构与能量差异。面向上述多掩码体系，建立与之匹配的分割学习框架，将多掩码信息引入损失加权与监督方式中，并在输出端加入“Core掩码为Hard掩码子集”的约束关系，实现对DRFM分量区域整体轮廓与高能核心区域的一致建模。仿真结果表明，相比仅依赖单一二值掩码的典型分割方案，所提出的多掩码标注与学习方法在DRFM分量像素级定位精度、核心区域提取能力以及掩码边界质量等方面均具有明显优势，为后续的DRFM分量抑制与目标恢复提供了更加精细且具有物理可解释性的空间先验。

This paper addresses the issue that the dense false target signals of digital radio frequency memory(DRFM) are difficult to be precisely located in the time-frequency domain.It focuses on how to provide interpretable pixel-level spatial priors for the subsequent suppression of DRFM components.Based on a unified simulation platform, the target, sea clutter and DRFM dense false target signals are jointly modeled.The complex-valued time-frequency diagrams obtained from the short-time Fourier transform(STFT) of the complete time-domain signals are taken as the main analysis objects.The time-frequency structures and energy distribution characteristics of the target, background and DRFM signals are depicted on the same time-frequency plane.On this basis, four types of mask supervision labels, namely soft mask, hard mask, core mask and valid mask, are constructed using the component decomposition information available in the simulation to monitor the hierarchical regional structure and energy differences of the DRFM components.For the above multi-mask system, a segmentation learning framework matching it is established.The multi-mask information is introduced into the loss weighting and supervision methods, and the constraint relationship of “the core mask is a subset of the hard mask” is added at the output end to achieve consistent modeling of the overall contour of the DRFM component region and the high-energy core region.Simulation results show that, compared with the typical segmentation scheme relying solely on a single binary mask, the proposed multi-mask annotation and learning method has significant advantages in pixel-level localization accuracy of DRFM component pixels, core region extraction ability and mask boundary quality.It provides a more refined and physically interpretable spatial prior for subsequent DRFM component suppression and target recovery.