Conditional GAN with Effective Attention for SAR-to-Optical Image Translation

Abstract

Synthetic aperture radar (SAR) is an effective observation technology, which is widely used in industry and agriculture. However, SAR images have speckle noise because of its imaging mechanism, so it is difficult to obtain useful information from them directly. Generative adversarial networks (GANs) have great performance in image translation with the development of deep learning, SAR images can be translated into optical images. However, due to the complex scene, low resolution and speckle noise, the generated images obtained by the existing methods are not satisfactory. In this paper, we propose a method based on conditional GAN (CGAN) for image translation from SAR images to optical images. We use the attention mechanism, which means that the network attaches importance to useful features and ignores unimportant ones. We apply discrete cosine transform (DCT) as loss function to extract the low frequency features in the image. Our experiments show that the quality of the images generated by our method is better than that of some famous methods.