This is the official repository for the paper "ProSR: Semantic-Prototype-Guided Discrete Modeling for Physically Consistent SAR Super-Resolution".
The curated dataset will be publicly available in this repository soon.
Synthetic Aperture Radar (SAR) super-resolution (SR) is a challenging task due to the severe scale ambiguity and speckle noise inherent in SAR imaging. Conventional methods often struggle to preserve the physically consistent local scattering structures, leading to blurry edges or structural distortion.
To address these issues, we propose ProSR, a novel framework that leverages Semantic-Prototype-Guided Discrete Modeling.
- Discrete Structural Constraints: We introduce discrete modeling to mitigate the SAR-specific LR ambiguity, preserving the physical consistency of reconstructed SAR images.
- Semantic-Prototype Guidance: By utilizing semantic prototypes, ProSR effectively guides the SR process to reconstruct high-fidelity structural details without severe distortion.
- Superior Performance: ProSR significantly improves the performance over the strongest baseline by 4.59%p in MSTAR ATR accuracy, demonstrating its practical effectiveness.
The experiments in our work are conducted on a curated, physically consistent SAR super-resolution dataset.
- The dataset is derived from the commercially available high-resolution SAR imagery provided by the Umbra Open Data Program.
- The pre-processed paired dataset (HR and LR) for training and evaluation will be released.
This dataset is a curated version of the data provided by the Umbra Open Data Program. The original raw Synthetic Aperture Radar (SAR) imagery was accessed via the Umbra Open Data on AWS Registry.
We faithfully acknowledge Umbra for making their high-resolution SAR data publicly available.
The curated dataset provided in this repository is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
According to the Umbra Open Data terms, any reuse or redistribution of this data must maintain this attribution.