【雷达与对抗】【2001】特征增强合成孔径雷达成像

本文为美国波士顿大学（作者：MUJDAT C ¨ ¸ ET˙ IN）的博士论文，共258页。

遥感图像已经在从天气预报到战场侦察的广泛任务中发挥着重要作用。最有前途的遥感技术之一是成像雷达，即合成孔径雷达（SAR）。SAR克服了光学相机的夜间限制，以及光学和红外成像仪的云层限制。在当前的系统中，使用极坐标格式算法等技术从采集的SAR数据中生成图像，这些图像由人类观察者解释。然而，预期的高数据率和新出现的SAR任务的时间紧迫性促使在从重建图像中提取信息时使用自动处理或决策技术。这种自动决策（例如目标识别）的成功与否取决于SAR图像对底层场景某些特征的显示程度。不幸的是，目前的SAR成像技术没有明确的手段来突出有助于自动判读的特征。此外，这些技术通常对减少的数据质量或数量不具有鲁棒性。

我们已经开发了一个具有数学基础和相关算法的特征增强SAR成像，以应对这些挑战。我们的框架是基于散射场的正则化重建，它将SAR观测过程的层析模型与感兴趣特征性质的先验信息相结合。我们通过各种非二次势函数证明了先验信息的包含。通过将半二次正则化方法推广到复值SAR问题中，实现了该框架下优化问题的高效、稳健数值求解。本文建立了一种基于识别特征的SAR成像技术定量评价方法。通过对大量真实和合成SAR图像的定性和定量分析，证明了特征增强成像的优点。这些优点包括提高分辨率、易于区域分割、旁瓣减少和斑点抑制，这些都是自动决策的重要属性。此外，我们通过在自动目标识别（ATR）系统上的分类实验，证明了特征增强SAR成像在提高自动决策性能方面的潜力。

Remotely sensed images have alreadyattained an important role in a wide spectrum of tasks ranging from weatherforecasting to battlefield reconnaissance. One of the most promising remotesensing technologies is the imaging radar, known as synthetic aperture radar (SAR).SAR overcomes the nighttime limitations of optical cameras, and the cloud-coverlimitations of both optical and infrared imagers. In current systems,techniques such as the polar format algorithm are used to form images from thecollected SAR data. These images are then interpreted by human observers.However, the anticipated high data rates and the time critical nature ofemerging SAR tasks motivate the use of automated processing or decision-makingtechniques in information extraction from the reconstructed images. The successof such automated decision-making (e.g. object recognition) depends on how wellSAR images exhibit certain features of the underlying scene. Unfortunately,current SAR image formation techniques have no explicit means to highlightfeatures useful for automatic interpretation. Furthermore, these techniques areusually not robust to reduced quality or quantity of data. We have developed amathematical foundation and associated algorithms for feature-enhanced SARimaging to address such challenges. Our framework is based on a regularizedreconstruction of the scattering field which combines a tomographic model ofthe SAR observation process with prior information regarding the nature of thefeatures of interest. We demonstrate the inclusion of prior information througha variety of non-quadratic potential functions. Efficient and robust numericalsolution of the optimization problems posed in our framework is achievedthrough novel extensions of half-quadratic regularization methods to thecomplex-valued SAR problem. We have established a methodology for quantitativeevaluation of a SAR image formation technique based on recognition-orientedfeatures. Through qualitative and quantitative analyses on large sets of realand synthetic SAR images, we have demonstrated the benefits provided byfeature-enhanced imaging. These benefits include increased resolution, ease ofregion segmentation, sidelobe reduction, and speckle suppression, which areimportant attributes for automated decision-making. Furthermore, we havedemonstrated the potential of feature-enhanced SAR imaging to improve automateddecision-making performance, through classification experiments on automatictarget recognition (ATR) systems.

1. 引言
2. 合成孔径雷达原理
3. 目前的SAR图像重建方法
4. 图像处理中的正则化方法
5. SAR成像中的正则化框架
6. 最优化问题的有效求解
7. 具有更一般势函数的图像重建
8. 其他变分公式的推广
9. 基于识别导向特征的评估

10.    特征增强图像识别测试
    

11.    特征增强重建在高分辨雷达中的应用
    

12.    结论与未来展望
    

附录A.1 离散二维微分算子
附录A.2 图像重建的目标函数梯度
附录A.3 半二次正则化的增广代价函数

更多精彩文章请关注公众号：