An Enhanced MIMO-Unet Model for CHASE High Resolution H\(\alpha \) Image Deblurring

Abstract

The Chinese H\(\alpha \) Solar Explorer (CHASE) satellite provided invaluable full-disk H\(\alpha \) spectral data, but images from its initial mission phase (before August 2022) suffered from significant defocus blur, limiting their scientific utility. This paper proposes an enhanced Multi-Input Multi-Output U-Net (MIMO-Unet) model to address this blind deconvolution challenge. Because sharp ground-truth counterparts are unavailable for real early-phase CHASE observations, we constructed a physically constrained dataset by applying a Zernike-polynomial-based defocus model to sharp CHASE/HIS observations acquired after the focusing condition became stable. The proposed model enhances the baseline MIMO-Unet by integrating Squeeze-and-Excitation (SE) blocks into the encoder for adaptive channel-wise feature recalibration and an Atrous Spatial Pyramid Pooling (ASPP) module into the feature fusion stage to capture multi-scale context. On the Zernike-simulated CHASE defocus dataset, the proposed method improves PSNR from 27.0169 dB to 33.8707 dB, corresponding to a 6.85 dB gain, and increases SSIM, FSIM, and the multi-fractal texture-based Perception Evaluation (PE) metric by 45.67%, 27.41%, and 62.96%, respectively, compared with the blurred inputs. Our model achieves the best PSNR (33.8707 dB) and SSIM (0.9282) among the compared deep learning methods, while being exceptionally lightweight with only 7.41M parameters. This work provides a physics-informed deep learning framework for restoring defocus-dominated degradation in CHASE H\(\alpha \) observations, with quantitative validation on Zernike-simulated defocus data and preliminary qualitative tests on real early-phase CHASE observations.