Wavefront Reconstruction Using Phase Diversity Based on the MobileViT Model

In adaptive Optics （AO） systems， wavefront aberrations are commonly represented by Zernike polynomials， whose coefficients describe distortions such as tilt， defocus， coma， and spherical aberration. Accurate estimation of Zernike coefficients is critical for high-quality imaging and effective wavefront correction. Due to the symmetry of the Fourier transform， Zernike coefficients with opposite signs can produce identical Point Spread Function （PSF） images， causing ambiguity in coefficient prediction. The Phase Diversity （PD） method mitigates this issue by simultaneously capturing focal and defocused intensity images， introducing additional phase information to improve reconstruction accuracy.Based on PD theory， this study proposes a MobileViT-based wavefront reconstruction method. The approach uses a pair of focal and defocused intensity images as input and automatically extracts image features to achieve non-iterative phase retrieval. The reconstruction performance of MobileViT is systematically evaluated under varying turbulence intensities and compared with CNN， ResNet-50， and EfficientNet-B0 models， providing a reference for practical applications.Experimental results demonstrate that MobileViT consistently outperforms the other models across all turbulence conditions. Although CNN has the fewest parameters and fastest inference， its reconstruction accuracy is the lowest. Under weak turbulence， all four models achieve reasonable wavefront reconstruction； as turbulence intensity increases， CNN's accuracy deteriorates， EfficientNet-B0 performs slightly worse than ResNet-50， and MobileViT maintains the highest accuracy， with the smallest residual RMS and largest Strehl Ratio （SR）， indicating stronger robustness under strong turbulence. Mean Absolute Error （MAE） comparisons of predicted versus true Zernike coefficients further show that MobileViT is least affected by turbulence， achieving superior precision and stability.Considering reconstruction accuracy， model size， training time， and inference speed， MobileViT achieves an optimal balance between lightweight design and high precision. This method demonstrates strong robustness under varying turbulence conditions， improving wavefront correction efficiency while reducing hardware dependency. These findings highlight the potential of combining deep learning with phase diversity for practical sensorless AO systems and suggest promising applications in fields such as astronomical imaging and laser communications. Future work will involve building an experimental platform to further validate the proposed approach.