稀疏正则驱动的超分辨SAR图像重建

Super-resolution SAR Image Reconstruction Based on Sparse Regularization Driven Mechanism

由于数据特性和散斑噪声影响,合成孔径雷达(Synthetic Aperture Radar,SAR)成像在传统数据驱动应用中存在瓶颈,特征可解释性弱。同时单纯模型驱动方法难以对复杂问题进行建模,灵活性差。针对上述问题,提出了一种稀疏正则驱动的超分辨SAR成像目标重建网络(SAR Target Reconstruction Network,SAR-TR-Net)。针对目标重建任务,从成像角度建模,将SAR复图像数据以及回波复数据引入模型驱动框架。然而SAR成像的信号处理过程包含矩阵乘法运算,并且成像模型中的观测矩阵也不能进行近似。为了适应成像模型的特殊性,设计了一种基于矩阵乘法约束的稀疏驱动框架,指导网络结构设计以实现数据模型双驱动的SAR图像重建。同时,在SAR-TR-Net卷积层采用交叉卷积结构保留原始SAR相位历程,提升了对复数据的特征提取能力。仿真和实测SAR数据验证结果表明,所提方法在目标杂波比、图像熵等量化指标上可以与成熟的模型驱动方法比拟,且在运行时间方面表现出色。

Due to the influence of data characteristics and speckle noise,there is a bottleneck in the application of conventional data-driven neural network in synthetic aperture radar(SAR)imagery,where the features to be extracted are hard to be interpreted.Moreover,it is difficult to formulate complicated problems with model-driven methods and the flexibility is poor.The novel super-resolution SAR target reconstruction network(SAR-TR-Net)is proposed based on sparse regularization to solve above problems.For the target reconstruction task,the modeling is carried out from the perspective of imagery,and the complex-valued SAR image data and echo data are introduced into the model-driven framework.However,the matrix multiplication operation is involved in the imagery signal processing process and the observation matrix of imagery model is not considered to be treated as an approximation.In order to adapt to the particularity of the imagery model,a sparse-driven framework based on matrix multiplication constraint is designed to guide the design of network structure for SAR imagery reconstruction based on data-driven and model-driven.Moreover,the cross-convolution structure is used to preserve the original SAR phase history in the convolutional layers,and the ability of complex-valued data feature extraction is improved.Simulation and measured SAR data verification demonstrate that the proposed method can be comparative with the mature model-driven method in terms of quantization indexes such as target clutter ratio and image entropy,and exhibites outstanding performance in terms of execution time.