基于迁移学习的SAR目标超分辨重建

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。

SAR target super-resolution based on transfer learning

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.