一种基于度量优化的深度网络SAR自聚焦方法

A Deep Network SAR Autofocusing Method Based on Metric Optimization

摘要: 相位误差是引起合成孔径雷达图像散焦的主要原因，受到目标场景与算法参数的影响，基于度量优化的传统自聚焦方法迭代次数多，计算量大。鉴于卷积神经网络在图像特征提取与特征分类方面的优越性，本文提出了一种基于度量优化的深度网络SAR自聚焦方法。首先，利用卷积神经网络的特征提取能力，将不同程度的离焦图像通过特征分类来实现离焦图像到相位误差系数的映射。其次，网络估计的相位误差系数被建模成相应的相位误差多项式，构成相应的误差向量来补偿离焦图像。通过使用度量函数作为损失函数进行网络训练，实现SAR离焦图像的误差更新和补偿。相比于传统自聚焦方法，基于深度网络的方法，训练完成后，不需要进行反复迭代，且不依赖于场景强散射点，具有更快的聚焦速度和稳定的聚焦性能。

Abstract: Phase error is the main cause of SAR image defocusing, which is affected by the target scene and algorithm parameters. The traditional autofocusing method based on metric optimization has many iterations and a large amount of computation. In view of the advantages of convolutional neural networks in image feature extraction and feature classification, in this paper, a autofocusing method for SAR in deep networks based on metric optimization is proposed. Firstly, the feature extraction capability of convolutional neural network is used to classify the defocusing images to realize the mapping from the defocusing image to the phase error coefficient. Secondly, the phase error coefficient estimated by the network is modeled into the corresponding phase error polynomial, and the corresponding error vector is formed to compensate the defocusing image. By using measurement function as loss function for network training, the error updating and compensation of SAR defocusing image are realized. Compared with traditional autofocusing methods, the method based on deep network does not need to carry out repeated iterations after training, and does not rely on strong scattering points in the scene, so it has faster focusing speed and stable focusing performance.