基于迁移学习的SAR目标超分辨重建  

SAR target super-resolution based on transfer learning

作　　者：徐舟[1] 曲长文[2] 何令琪 

机构地区：[1]电子工程学院雷达对抗系,合肥230000 [2]海军航空工程学院电子信息工程系,烟台264001 [3]北京大学软件与微电子学院,北京100871

出　　处：《航空学报》2015年第6期1940-1952,共13页Acta Aeronautica et Astronautica Sinica

基　　金：国家自然科学基金(61102166);山东省优秀中青年科学家科研奖励基金(BS2013DX003)~~

摘　　要：针对合成孔径雷达（SAR）目标超分辨重建问题,提出了一种基于迁移学习的超分辨方法。在光学图像梯度域中联合训练超完备字典与稀疏编码映射,利用半耦合字典联系SAR图像与光学图像,寻找SAR图像在半耦合字典下的稀疏编码,并在高分辨率字典下完成重建。结合SAR图像的先验信息,使用正则化方法对SAR目标进行特征增强。所提方法在TerraSAR-X数据和MSTAR数据上进行了仿真实验,重建结果表明,相比目前的插值方法和稀疏表示方法,所提方法空间分辨率可提高0.5-1.5个像素。正则化增强结果表明,引入稀疏先验的正则化增强能够进一步提高空间分辨率并抑制杂波比,最后分析了正则化参数的选取对图像质量的影响。Based on transfer learning, a method for synthetic aperture radar（SAR） target super-resolution reconstruction is proposed in this paper. A semi-coupled dictionary is jointly trained in the gradient domain of optical image. By utilizing the relationship revealed by semi-coupled dictionary, the sparse codes of SAR image are obtained. Then the image is reconstructed in the high resolution dictionary. Based on some prior knowledge of SAR image, the regularization method is also used in order to enhance the target feature. Several simulation experiments are conducted based on TerraSAR-X and MSTAR data, and the reconstructed results show that the spatial resolution obtained by the proposed method is 0.5-1.5 pixels higher compared to the current interpolation method as well as the sparse representation method. Regularization enhancement results show that it can further improve the spatial resolution and suppress clutters by introducing the sparse prior. Finally, the influences on the spatial resolution and target structure of the reconstruction image caused by regularization parameter are analyzed qualitatively.

关 键 词：合成孔径雷达 超分辨 迁移学习 半耦合字典 稀疏表示 

分 类 号：V557.3[航空宇航科学与技术—人机与环境工程] TP751.1[自动化与计算机技术—检测技术与自动化装置]