Improved cGAN for SAR-to-Optical Image Translation

Synthetic aperture radar (SAR) can be used for all-day and all-weather Earth observation, but it has the disadvantages of speckle noise and geometric distortion, which are not conducive to human eye recognition. In order to enhance the visual effect of SAR images, this paper proposes an improved cGAN(Conditional Generative Adversarial Network) method to achieve the translation of SAR images to optical images. Firstly, the generator uses U-Net structure to combine global features with local features, which improves the details of the generated image. Secondly, the discriminator uses PatchGAN structure to extract and characterize the local image features, and finely distinguish each part of the image. Finally, SSIM and PSNR loss functions are added to improve the restoration degree of the generated image. In the experiment on SEN1-2 dataset, our method surpasses the basic cGAN and pix2pix. The translated image retains the key content of SAR image, and also has the style of optical image.