HSFMamba: Hierarchical Selective Fusion Mamba Network for Optics-Guided Joint Super-Resolution and Denoising of Noise-Corrupted SAR Images
Year: 2025
Type: article
Abstract: Synthetic Aperture Radar (SAR) image interpretation has attracted widespread attention in remote sensing applications. However, the performance of existing SAR image interpretation methods is severely hindered by the inherent limitations of SAR imaging mechanisms, such as speckle noise and low resolution. With the continuous advancement of remote sensing technology, it has become increasingly feasible to simultaneously acquire optical and SAR images. Given the rich details in optical images, it is crucial to exploit this valuable information to guide the quality enhancement of SAR images, thereby significantly improving their performance for practical applications. In this work, we propose a novel Hierarchical Selective Fusion Mamba Network (HSFMamba) for Optics-Guided joint Super-Resolution (SR) and Denoising of SAR images, which simultaneously addresses the resolution limitations and noise corruption in a unified framework. HSFMamba leverages the long-range modeling capability of the State Space Model (SSM) with linear complexity and incorporates optical images through two progressive cross-selection scan mechanisms to perform high-quality reconstruction of SAR images corrupted by speckle noise. Specifically, we design a Cross-Modal Feature Selection Module (CFSM) that dynamically identifies significant feature representations in optical images, thereby progressively extracting key information from the optical images. Additionally, to effectively leverage optical image details while mitigating SAR speckle noise, we develop a Frequency-Spatial Adaptive Aggregation Module (FSAM) aimed at better restoring image details, effectively enhancing critical high-frequency information. In addition, to facilitate research in this domain, we construct a well-aligned and high-resolution dataset for optics-guided joint SAR image SR and denoising, consisting of 3,200 pairs of optical and SAR images, with a total of 25,600 image pairs across eight degradation modes. Extensive experiments demonstrate that HSFMamba effectively utilizes optical image information to improve SAR image quality, outperforming several state-of-the-art methods. The dataset is available athttps://github.com/mmic-lcl/Datasets-and-benchmark-code.
Authors: Zhicheng Zhao, Qing Gao, Jinquan Yan, Chenglong Li, Jin Tang
Institution: Anhui University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W4411446463