Spaceborne SAR Ship Data Expansion Based on GAN

In response to the lack of high-quality and sufficient synthetic aperture radar (SAR) data that can be paired with light ship images in practice, this paper proposes a method to convert optical ship images into SAR images based on GAN networks. In order to improve the coordination of generated images and stabilize the training process, this paper establishes a generator with a self-attention mechanism combined with spectral normalization operations. Secondly, the maximum spatial perturbation consistency loss is introduced to reduce the spatial changes between optical images and SAR images caused by changes in ship size and background noise. Finally, in order to speed up the convergence speed and stability of the model, a relative discriminator is introduced. Experiments conducted on the HRSID and LEVIR remote sensing ship data sets show that the FID (Fréchet Inception Distance) of the SAR ship images generated by this model is reduced to 57.25. At the same time, compared with the current optimal algorithm, the peak signal-to-noise ratio is increased by 5.89% on average. The structural similarity loss is increased by 18.75% on average, and the mean square error is reduced by 13.44% on average.