An ISAR Image Generation Method with New Azimuth Based on Generative Adversarial Network

Due to ISAR images' strong pose sensitivity, they have significant differences under different azimuth conditions. However, obtaining ISAR images of the target's full pose in practical applications is difficult, which can adversely affect downstream tasks such as deception jamming. Therefore, it is necessary to use limited azimuth ISAR images to generate ISAR images with missing azimuth angles and supplement the missing data. This paper is inspired by the Conditional Generative Adversarial Networks (CGAN) structure and designs a novel network for the multi-azimuth rotation task of ISAR images to meet the needs of multi-angle sample expansion. We use azimuth information as conditional input to train the network and establish a mapping between azimuth information and the generated target pose. In addition, to solve the problem of pattern collapse, we use ISAR images and their feature maps as inputs to make the generated images more in line with the physical features of the target. Meanwhile, a comprehensive loss function was designed that combines adversarial loss, reconstruction loss, and angle loss to ensure that the generated image is visually and structurally closer to the actual ISAR image. The experimental results show that the difference between the generated ISAR image and the real label image is slight when our network's rotation angle is small. Especially when the rotation angle is controlled within 10°, the network can generate high-quality images that are highly similar to real ISAR images, and the network's strong generalization ability provides strong support for achieving effective deception jamming.