SAR-ASGOUT: Adversarial Speckle Generation and Offline Unpaired Training for Unsupervised SAR Despeckling

The presence of inherent speckle noise poses considerable challenges to the intelligent application of synthetic aperture radar (SAR) images. Existing deep learning-based despeckling methods have evolved into two primary categories: supervised and unsupervised approaches. Supervised methods are often impractical because of the lack of clean reference SAR images, while unsupervised methods frequently underperform, largely due to the lack of effective prior information constraints. Recent advancements in unpaired unsupervised methods have attempted to mitigate these issues, yet they still grapple with variability in the generated speckle noise quality and the complexity of model tuning. To overcome these limitations, we propose an unsupervised method based on adversarial speckle generation and offline unpaired training for SAR despeckling (named SAR-ASGOUT). Our approach introduces two key innovations: preliminary despeckling reference filtering (PDRF) and adversarial speckle distribution generating (ASDG), which generate speckle noise that accurately reflects the true statistical properties. In addition, we develop a statistical and spatial consistency speckle database (SSC-SD), enhanced with a speckle preprocessing module designed to fine-tune the statistical attributes of the speckle and preserve its spatial correlations. The unpaired speckled-clean image pairs in SSC-SD facilitate stable offline training of the despeckling network, leveraging the strengths of both supervised and unsupervised methodologies. By integrating adversarial speckle generation with offline unpaired training, SAR-ASGOUT effectively overcomes the limitations of existing methods. Extensive experimental evaluations demonstrate that SAR-ASGOUT surpasses current state-of-the-art methods in despeckling performance while also excelling in preserving fine details in real SAR images, offering a robust solution for SAR image enhancement.