基于GAN的ISAR图像语义分割方法

逆合成孔径雷达（ISAR）成像技术能够对空间目标进行远距离成像，刻画目标的外形、结构和尺寸等信息。ISAR图像语义分割能够获取目标的感兴趣区域，是ISAR图像解译的重要技术支撑，具有非常重要的研究价值。由于ISAR图像表征性较差，图像中散射点的不连续和强散射点存在的旁瓣效应使得人工精准标注十分困难，基于交叉熵损失的传统深度学习语义分割方法在语义标注不精准情况下无法保证分割性能的稳健。针对这一问题，提出了一种基于生成对抗网络（GAN）的ISAR图像语义分割方法，采用对抗学习思想学习ISAR图像分布到其语义分割图像分布的映射关系，同时通过构建分割图像的局部信息和全局信息来保证语义分割的精度。基于仿真卫星目标ISAR图像数据集的实验结果证明，本文方法能够取得较好的语义分割结果，且在语义标注不够精准的情况下模型更稳健。

ISAR Image Semantic Segmentation Based on GAN

Inverse synthetic aperture radar (ISAR) imaging technology can perform long-range imaging of space targets, depict the information such as the shape, structure and size of the target. Semantic segmentation of ISAR images can obtain the region of interest of the target, which is an important technical support for ISAR image interpretation. Semantic segmentation of ISAR images has very important research value. Due to the poor representation of ISAR images, the discontinuity of the scattering points in the image and the sidelobe effects of strong scattering points make it very difficult to label accurately. Traditional deep semantic segmentation methods based on cross-entropy loss cannot guarantee the robustness of segmentation performance when labels are not accurate. In this paper, an ISAR image semantic segmentation method based on generative adversarial networks (GANs) is proposed. Adversarial learning ideas are used to learn the mapping relationship from the distribution of ISAR images to the distribution of corresponding semantic segmentation images. At the same time, the accuracy of semantic segmentation is ensured by constructing the local information and global information of the segmented image. The experimental results based on the simulated satellite target ISAR image dataset prove that the proposed method can obtain better semantic segmentation results, and the model is more robust when the semantic labels are not enough accurate.