Synthetic Aperture Radar (SAR) is an advanced remote sensing tool that utilizes microwave waves to produce high-resolution images of the Earth's surface. This technology is capable of collecting precise data under various weather conditions. Due to the use of coherent waves, SAR images encounter a phenomenon known as Speckle noise. Removing Speckle noise is a crucial step in processing the amplitude or intensity images generated by coherent systems like SAR. Since Speckle is caused by the specific characteristics of the targets in the image, accurately identifying these characteristics is essential for effectively estimating and reducing the noise. In this thesis, traditional methods for Speckle noise reduction are first reviewed. These methods are generally categorized into three groups: Bayesian methods in the Spatial domain, Bayesian methods in the Transform domain, and non-Bayesian methods. Then, novel Deep learning-based methods, including Supervised and Self-supervised learning approaches for speckle noise reduction, are introduced. In supervised learning, two approaches—direct deep learning and model-based deep learning—are examined. This thesis proposes a novel supervised approach based on Wasserstein Generative Adversarial Networks (Wasserstein GANs) for Speckle noise reduction in SAR images. The proposed method utilizes two discriminator networks, the Sobel filter to preserve image details, and diverse loss functions. Quantitative and experimental results from eva‎luation on simulated images show that the PSNR for the proposed network is about 9 dB higher on standard test images compared to traditional methods. Additionally, on real images, the ENL value is 0.6 higher than that of the best traditional method. Therefore, while the proposed method outperforms traditional methods in reducing speckle noise, it can also effectively preserve the structural details of the image.

محمود مدرس هاشمي

مهران صفاياني

محمدرضا احمدزاده , احسان يزديان