A Wavelet-Enhanced and Physically-Informed Network for Target-Level SAR-to-Optical Image Translation

Synthetic aperture radar (SAR) imaging technology has been widely used in the field of remote sensing. However, due to the electromagnetic wave imaging principle of SAR, SAR images often have speckle noise and image distortion, which brings greater complexity to image interpretation and downstream tasks. Therefore, converting SAR images with poor interpretation into optical images with rich spectral information is helpful for the analysis of SAR images and promotes the implementation of downstream tasks. Based on the unique imaging mechanism of SAR images, we pay more attention to the frequency domain information of the image and the physical information of the target, propose a new generative adversarial network (GAN) to construct a conversion network from SAR to optical images. Through wavelet transform, a frequency-aware model is constructed, so that the conversion model can retain more contour and structural information and improve the image conversion quality. In addition, we establish a scattering center model of the target based on the electromagnetic characteristics of the target, and draw on the idea of contrastive learning to guide the network to extract the physical information of the target more efficiently. The experimental results on the dataset show that our proposed method outperforms many advanced methods in multiple metrics in the generation field. In addition, our model can also be used to assist downstream tasks of SAR images. The three-dimensional reconstruction task of SAR targets assisted by our proposed translation model is also improved compared with our previous method.