Highly accurate centroid tracking for real-time event-based Shack–Hartmann wavefront sensing

Shack–Hartmann wavefront sensors (SHWFSs) measure wavefront aberrations for use in adaptive optics compensation, which relies on low-latency tracking of an array of lenslet focus spots. Prior field experiments have shown that event-based cameras (EBCs) can effectively replace conventional frame-based cameras (FBCs) in SHWFSs by reporting spot movements as log-intensity changes with microsecond time resolution and low latency. While state-of-the-art EBC SHWFS methodologies employ convolutional neural networks that incur substantial computational cost, this work targets both higher estimation accuracy and markedly reduced complexity. Specifically, we propose a recurrent neural network (RNN) called the Shack–Hartmann event-based recurrent neural network (SHEBRNN) that takes as input a stream of event data—pixel position, polarity, and time between events—and predicts the spot centroid position. The network was trained with data from a custom SHWFS hardware capturing FBC and EBC simultaneously, using the spot centroid positions computed from the FBC as pseudo-ground truth to train/test the event-RNN-based centroid position estimation method in an unsupervised manner. Our network improves the slope estimation accuracy by 1 µrad over the state of the art, and the achieved wavefront reconstruction Strehl ratio exceeds 91% while gaining substantial computational efficiency. By normalizing over the microlens’s focal length, we also achieve stable performance over various optical configurations.