"SAR Image Denoising and Colorization Dataset"

"The dataset under consideration is a set of 10,000 paired Synthetic Aperture Radar (SAR) and optical-style reference images created to aid in the research of SAR image denoising and perceptual colorization. The dataset is made to imitate real Earth-surface patterns, such as agricultural land, city regions, vegetations, water features and barren area while maintaining geographical anonymity and complete syntheticity to make it license-free to public distribution.The data is organized in a way that allows two-stage learning pipeline. During the initial step, clean grayscale optical-like images are offered as luminance (L) references to speckle noise reduction in SAR images. The second stage involves the regression targets of SAR-to-optical colorization using corresponding chrominance (a, b) in the CIE Lab color space. The supervision is done using pixel-level regression targets in place of discrete semantic class labels.All the images are rescaled to 224 224 pixels, and they are kept without logos, watermarks, comments, and artificial interventions. The dataset is accompanied by a structured metadata file that records the identifiers of the images, the data splits, the preprocessing stages, and the information on the specific use of the task.The data can be used to facilitate research and education in the areas of SAR image denoising, SAR-to-optical translation, image colorization and evaluation of two-stage deep learning architectures in remote sensing and perceptual image optimization schemes."