Optical-ISAR Image Translation via Denoising Diffusion Implicit Model
Xinyi Wang, Huaizhang Liao, Zhixiong Yang, Tao Han, Jingyuan Xia · 2023
Sufficient training data play a crucial role in deep-learning techniques, however, available training samples are not always accessible for specific targets, such as satellites. Producing synthetic data through generative model is a major solution for the lack of training data. In this paper, we propose a Denoising Diffusion Implicit Model (DDIM) to generate the Inverse Synthetic Aperture Radar (ISAR) satellite images from the synthesized optical counterparts. Efficient and abbreviated feature extraction method is proposed in optical-ISAR image translation. A domain-cross generative model is then established with respect to the generative task and feature learning modules obtained from the translation model. Extensive experiments validate the feasibility and performance of the proposed approach, which outperforms the existing generative models such as CycleGAN in terms of inception score and structural similarity index.