Speckle-Driven Unsupervised Despeckling for SAR Images

Speckle noise in Synthetic Aperture Radar (SAR) images significantly degrades image quality and complicates image interpretation. Therefore, reducing speckle noise in SAR images is a crucial step. Recently, deep learning-based methods have demonstrated impressive performance in SAR image despeckling. However, due to the unavailability of clean SAR images, supervised despeckling models trained on synthetic noisy images often suffer from poor generalization. To address this challenge, we propose a speckle-driven unsupervised despeckling network named SDUDNet. This network employs a generative adversarial network (GAN) strategy to overcome the lack of paired data by utilizing only unpaired clean and noisy training images. Specifically, the generator in SDUDNet learns the noise distribution in real SAR images and transforms clean images into pseudo-SAR images without any assumptions or priors. Additionally, we introduce a global-local discriminator (GLD) to finely discern subtle features in the generated images. Furthermore, an attention network is integrated to help the discriminator focus on critical details within the images. Experiments conducted on both synthetic and real SAR images demonstrate that SDUDNet achieves state-of-the-art results in terms of visual quality and objective metrics. The source code of our proposed method is available at https://github.com/BFY-official/SDUDNet.