Abstract
The rotating synthetic aperture (RSA) optical imaging system employs a rectangular primary mirror for detection. During the imaging process, the primary mirror rotates around the center to achieve the aperture equivalent to the long side of the rectangle at different rotation angles. As a result, the system's point spread function (PSF) changes over time, causing periodic time-varying characteristics in the acquired images' resolution. Moreover, due to the rectangular primary mirror, the images obtained by the RSA system are spatially asymmetric, with a lower resolution in the short side's direction than in the long side's direction. This paper characterizes the imaging quality degradation mechanism of the RSA system and the time-space evolution law of the imaging process. Through a semi-physical experimental platform, we quantify the impact on image degradation and evaluate the limitations of generic deep learning-based super-resolution methods in addressing these unique artifacts.
Introduction
In high-orbit remote sensing, the demand for higher resolution necessitates larger primary mirrors. However, traditional monolithic circular mirrors are heavy and difficult to launch. The Rotating Synthetic Aperture (RSA) system offers a lightweight alternative by using a slim rectangular mirror with a high aspect ratio. By rotating this mirror, the system can theoretically synthesize the resolution of a large circular aperture over time.
However, this design introduces a complex "spatially asymmetric and temporally periodic" blur. A rectangular aperture produces a PSF that is narrow in the direction of the long side (high resolution) and wide in the direction of the short side (low resolution). As the mirror spins, this elliptical blur rotates, creating images with resolution that varies both across the frame and over time.
Theoretical Modeling and Experimental Setup
The researchers first derived the mathematical PSF for a rotating rectangular pupil. The analysis confirmed that the system’s energy distribution is roughly elliptical and rotates in synchronization with the mirror's physical orientation.
To verify these characteristics, a semi-physical simulation platform was constructed. Rather than relying solely on digital simulation, the team used a high-quality circular mirror and placed rotating rectangular mask plates in front of it to simulate various mirror geometries and aspect ratios. High-resolution remote sensing images from the WorldView-3 satellite—covering urban infrastructure and natural landscapes—were passed through this rig to create a specialized dataset of RSA-degraded image pairs.
Evaluation of Super-Resolution Methods
The study tested five state-of-the-art deep learning Super-Resolution (SR) networks: SRGAN, EDSR, SRMD, FSSR, and Real-ESRGAN. The goal was to determine if generic models, typically trained on isotropic (uniform) blur, could compensate for the extreme directional information loss inherent in RSA systems.
Key Findings
- The Directional Information Gap: Resolution in the direction of the mirror's short side is dramatically lower than the long side. This creates a "striped" or "smeared" effect that changes orientation as the mirror rotates.
- Texture Sensitivity: The system impacts complex textures (e.g., airports or forests) more severely than uniform scenes (e.g., open sea). PSNR and SSIM scores were consistently 5–10% lower for high-texture scenes.
- Generic AI Limitations: While models like SRMD provided a quantitative boost (up to ~6 dB in PSNR), they failed to restore fine details in the "low-resolution" direction of the ellipse. Because these models are designed for uniform blur, they struggle with the extreme asymmetry of high-aspect-ratio mirrors.
- Aspect Ratio Threshold: When the mirror's length-width ratio exceeds 7, image quality drops below the threshold required for reliable interpretation, even after AI enhancement.
Conclusion
The RSA system presents a unique challenge for image restoration due to its time-varying, asymmetric PSF. Current generic SR models treat RSA blur as a standard artifact rather than a mathematically predictable projection, leading to suboptimal reconstruction of directional details.
Future work should focus on physics-informed deep learning models. By incorporating the mirror's rotation angle and aspect ratio as priors, multi-frame reconstruction algorithms could potentially fuse a sequence of rotating images to synthesize a single, high-resolution circular-equivalent image, overcoming the limitations of single-image processing.
Summary updated