CV-SAGAN: Complex-valued Self-attention GAN on Radar Clutter Suppression and Target Detection
Yuanhang Wu, Chenyu Zhang, Yiru Lin, Xiaoxi Ma, Wei Yi
Abstract
Traditional clutter suppression and target detection methods have limitations in that they must satisfy specific statistical models. In this paper, we propose a unified deep learning model for complex-valued self-attention generative adversarial networks (CV-SAGAN) for clutter suppression and target detection. In the complex-valued framework, we first use a generator module to learn the clutter distribution and perform clutter suppression. Then, a self-attention module is used for the first time to perform corrective detection of sparse targets. Finally, a discriminator is used to judge between the real data and the network output results, improving the robustness of the model. We verified that the CV-SAGAN model has a better detection rate and robustness than the conventional cell-average constant false alarm rate (CA-CFAR), real-valued GAN, and real-valued SAGAN on three clutter distributions and achieved better detection results on the publicly available IPIX real-world dataset.