Fig. 1. Downloaded target celestial images from Hubble official website

Fig. 2. Degraded images of simulated target celestial bodies subjected to atmospheric turbulence with different intensities. (a) Degraded images when k=0.001; (b) degraded images when k=0.0025; (c) degraded images when k=0.005

Fig. 3. Generative adversarial network

Fig. 4. Multi-scale feature fusion

Fig. 5. Topology diagram of generated network architecture

Fig. 6. Topology structure diagram of BmffGAN overall network^[19]

Fig. 7. Process of BmffGAN training. (a) Curve of loss function changing with epoch; (b) curve of PSNR changing with epoch

Fig. 8. Comparison of restoration effects of different algorithms on simulated atmospheric turbulence images under k=0.005. (a) Degraded images when the additive noise mean is 0, and the variance is 0.001; (b) SGL algorithm; (c) CLEAR algorithm; (d) IBD algorithm; (e) DNCNN algorithm; (f) BmffGAN

Fig. 9. Comparison of restoration effects of different algorithms on simulated atmospheric turbulence images under k=0.0025. (a) Degraded images when the additive noise mean is 0, and the variance is 0.001; (b) SGL algorithm; (c) CLEAR algorithm; (d) IBD algorithm; (e) DNCNN algorithm; (f) BmffGAN

Fig. 10. Comparison of restoration effects of different algorithms on simulated atmospheric turbulence images under k=0.001. (a) Degraded images when the additive noise mean is 0, and the variance is 0.001; (b) SGL algorithm; (c) CLEAR algorithm; (d) IBD algorithm; (e) DNCNN algorithm; (f) BmffGAN

Fig. 11. Evaluation indexes of different algorithms for restoration of simulated atmospheric turbulence images with different intensities (average value). (a) PSNR; (b) SSIM; (c) GMSD; (d) recovery time

Fig. 12. Munin ground-based telescope and star chart software

Fig. 13. Comparison experiment for restoring the ISS images affected by real turbulence. (a) ISS images affected by real turbulence; (b) SGL algorithm; (c) CLEAR algorithm; (d) IBD algorithm; (e) DNCNN algorithm; (f) BmffGAN