My Contribution
1.
Designed an object-independent error metric and its gradient to implement parallel framework.
•
Our algorithm is designed with intrinsic parallelism, allowing not only for optimized GPU acceleration but also for enhanced performance on CPU via multi-threading.
2.
Reduced computational burden by using approximate model for point spread function (PSF).
3.
Validation of real-time feasibility using real-world data based on the Kolmogorov model.
Schematic diagram of image-based wavefront sensing based on phase diversity method
In Details
Mathematical modeling for image-based wavefront sensing
1.
Image intensity distribution
•
Convolution between object image and PSF
2.
Comparison of different PSF models
•
PSF approximation model based on Taylor-series expansion
•
PSF and its gradient in spatial frequency domain
◦
Fourier-Optics PSF (FOPSF) : FFT computation → Computational burden ↑
◦
Approximate PSF (APPSF) : Element-wise multiplication → Computational burden ↓
•
PSF 2D image comparison
3.
Derivation of object-independent error metric
•
An ideal object image is derived by minimizing the intensity difference between measured and predicted images from the optical system
Gradient-based optimization algorithm
1.
Derivation of the gradient for the object-independent error metric
•
Establish an intrinsic parallel structure by separating each Zernike modes
2.
Comparison of local optimization algorithms
•
ADAML → more stable & faster
3.
Overall procedure
Numerical Experiments
1.
Wavefront sensing
•
GPU acceleration effect : GPU computation time changed very little for the number of states.
2.
Wavefront correction
•
RMSE and computation time for the number of states
3.
Image correction effect
•
No discernible difference in measurement performance between global optimization and local optimization
