Convolutional neural network for improved event-based Shack-Hartmann wavefront reconstruction

Area of Science:

• Optics and Photonics
• Machine Learning
• Adaptive Optics

Background:

• Shack-Hartmann wavefront sensing (SHWFS) traditionally uses frame-based cameras for aberration measurement.
• Conventional methods are limited by fixed sampling rates and inefficient pixel usage.
• Event-based cameras offer asynchronous, high-speed data acquisition for SHWFS.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) for real-time, accurate spot centroid estimation in event-based SHWFS.
• To evaluate the performance of the proposed EBWFNet in real-world scenarios.
• To compare the EBWFNet against state-of-the-art event-based SHWFS techniques.

Main Methods:

• Development of a custom SHWFS hardware with synchronized frame- and event-based cameras.
• Implementation of an event-based wavefront network (EBWFNet) using CNN architecture.
• Unsupervised training and testing of the EBWFNet utilizing frame-based camera data.
• Field testing and ablation studies to assess performance and component impact.

Main Results:

• The EBWFNet achieved highly accurate, sub-pixel spot centroid estimation in real-world conditions.
• Demonstrated substantial improvement over existing state-of-the-art event-based SHWFS methods.
• An unoptimized MATLAB implementation achieved speeds exceeding 800 Hz on a single GPU.

Conclusions:

• The proposed EBWFNet significantly enhances the accuracy and speed of event-based SHWFS.
• Event-based cameras coupled with CNNs represent a promising advancement for adaptive optics.
• The EBWFNet offers a robust and efficient solution for real-time wavefront aberration measurement.