MD-Glow: Multi-task Despeckling Glow for SAR Image Enhancement

Shunsuke Takao; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 536-543

Synthetic Aperture Radar (SAR) is a widespread technology to measure data in earth observation or environmental monitoring however severe speckle noise degrades SAR image quality which hinders in-depth image analysis. Speckle noise in a SAR image is mitigated by deep learning methods although they suffer from a domain gap between training and testing data to pose problems such as oversmoothing in the recovered images. It is also challenging for practical applications to enhance satellite imagery by a single model often having multiple scales and resolutions measured by different bands or sensing devices. In this paper we propose a novel scheme of Multi-task Despeckling Glow dubbed as MD-Glow for SAR image enhancement. Our deep model suppresses speckle noise (despeckling task) while simultaneously enhancing image resolution (super-resolution task) in a unified scheme to effectively address the above problems for despeckling. We conduct numerical experiments on the despeckling task utilizing datasets composed of synthetic and real SAR images to demonstrate that the proposed MD-Glow produces competitive results in comparison to other despeckling methods.