Context2Context: A Zero-Shot SAR Image Speckle Filter

Advanced speckle filters can well preserve textural features and fine details in synthetic aperture radar (SAR) images, facilitating subsequent interpretations and applications. In recent years, a multitude of deep learning-based despeckling methods have been developed. However, most of them belong to supervised learning fashion trained on simulated datasets which may affected by the domain gap between simulated and real measured SAR data. Meanwhile, two hinds of self-supervised learning methods have been reported. The first hind mainly requires multiple SAR images of the same region for training, while these images are costly or not always available. The second hind mainly constructs lower resolution image pairs for training with sublook decomposition, which may degrade final speckle reduction performance. In order to fulfill aforementioned gaps laying in self-supervised learning methods, a zero-shot SAR image speckle filter is proposed. Its innovation contributions are three-fold. Firstly, the theoretical principle that model training with speckle-speckle image pairs is equivalent to speckle-clean image pairs is further manifested. Secondly, a similar pixel bank scheme is established which can collect sufficient similar pixels and construct plenty of independent training samples based on contextual information without necessity of multiple observation images and image resolution degradation. Finally, a zero-shot speckle filter framework is developed and well trained. Extensive experiments are performed using various SAR datasets in terms of diverse frequency bands and spatial resolutions from Radarsat-2, MiniSAR, and ALOS-2. Comparisons with state-of-the-art methods demonstrate the superiority of the proposed method. Experimental results show that the proposed Context2Context speckle filter can effectively smooth homogeneous regions while preserving target characteristics from both visual quantitative examinations effects.