A new blind single-frame restoration framework, ASTRA-SR, significantly improves the super-resolution of astronomical images by addressing common observational challenges. The framework, detailed in a paper on arXiv cs.CV, was trained using a physics-grounded synthetic dataset to combat atmospheric turbulence, sensor noise, and limited sampling that typically degrade ground-based planetary imaging. This approach offers a 0.49 dB foreground PSNR gain, indicating a substantial improvement in image quality.
Ground-based telescopes contend with atmospheric turbulence, which causes blurring and image jitter due to temperature and pressure fluctuations in Earth's atmosphere. Traditional methods, such as "lucky imaging," capture numerous frames and select the sharpest ones, but this can be data-intensive and prone to noise accumulation. ASTRA-SR bypasses these limitations by employing a single-frame restoration process.
The development of ASTRA-SR involved creating a specialized synthetic dataset. This dataset uses high-dynamic-range raw observations from spacecraft as clean source images. Low-resolution input images are then synthesized by incorporating measured turbulence strengths, propagated moving phase screens, exposure-averaged spatially varying point spread functions (PSFs), and realistic sensor noise. This physics-based synthesis allows ASTRA-SR to learn how to effectively reverse these degradation effects.
The ASTRA-SR framework operates in two main stages. First, it estimates a noise-suppressed image that still retains blur. Following this, it restores the spatial structure through multi-scale processing and reconstructs high-resolution details using a serial spatial-amplitude refinement technique. This method allows for the recovery of fine details often lost in turbulent atmospheric conditions.
Other research efforts in astronomical super-resolution also focus on creating improved datasets and models. The STAR dataset, for instance, provides 54,738 flux-consistent star field image pairs, combining Hubble Space Telescope observations with physically faithful low-resolution counterparts. This dataset aims to address limitations in previous astronomical super-resolution datasets, such as flux inconsistency and insufficient data diversity. Similarly, AstroDiff, another generative restoration method, leverages diffusion models to mitigate atmospheric turbulence in astronomical images, showing improved perceptual quality and structural fidelity.
The advancements in frameworks like ASTRA-SR and AstroDiff, along with specialized datasets, contribute to the ongoing effort to enhance the quality of astronomical observations. Improved image resolution is crucial for detecting distant celestial objects and conducting precise structural analysis. The ability to generate clearer images from ground-based telescopes can also lead to more cost-effective high-resolution capture.