Lightweight convolutional neural network for wavefront reconstruction via a Shack-Hartmann sensor with spatially...

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Astronomy Instrumentation

Background:

• Wavefront sensing is crucial for adaptive optics systems, especially in astronomy.
• Traditional Hartmann sensors struggle with weak beacon light and strong atmospheric turbulence.
• Need for faster, more efficient wavefront detection methods.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) for rapid wavefront phase prediction.
• To address limitations of Hartmann sensors under weak light and strong turbulence.
• To improve the speed and reduce the computational complexity of wavefront sensing.

Main Methods:

• Designed SIR-Net, a lightweight CNN utilizing parallel processing of sub-spots.
• Employed a special downsampled microlens Hartmann design for improved Signal-to-Noise Ratio (SNR).
• Applied the average magnitude of variation method to reduce model complexity and computational load.

Main Results:

• Achieved wavefront root-mean-square residuals of 0.059λ under atmospheric turbulence (d/r₀≈3).
• Demonstrated a fast inference time of 0.191 ms.
• Developed a compact model with a size of 1.772 MB.

Conclusions:

• SIR-Net effectively predicts wavefront phase from Hartmann images, even with weak light.
• The proposed method significantly reduces inference time and model size.
• SIR-Net offers a viable solution for rapid wavefront detection in challenging environments.