Task
Participants are required to reconstruct detailed astronomical images by removing turbulence effects using multispectral full-disk video data with and without turbulence.
Data
This dataset contains 2000 groups of data (1000 real and 1000 simulated), split into 50% for training, 5% for validation, and 45% for final evaluation. The real data pairs include ground-truth and turbulence-affected images, while the simulated data pairs include ground-truth and turbulence-simulated images. The overall dataset consists of 70% partial solar images and 30% full solar disk images, with dimensions of 500 × 500 × 21 × 10. The dataset is stored in .mat format.
Metrics
PSNR, SSIM, and LPIPS are employed for the quantitative evaluation of the simulated dataset. For ground-based turbulence-degraded images without reference, we use Fourier ring correlation to assess the detail restoration in re- constructed images and compute the reconstruction stability across frames using cross-correlation matrices and inter-frame slices.