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

This study presents SIR-Net, a lightweight convolutional neural network designed to predict wavefront phase from images of a Hartmann with special downsampled microlens. The Hartmann is structured to achieve a suitable SNR when the light energy of the beacon is weak. The network extracts features from each sub-spot in parallel processing, which speeds up the inference rate. The average magnitude of variation method is applied to remove unimportant parameters and reduce computational complexity. Both numerical simulation and experimental results prove the feasibility of the proposed approach. The results indicate that under atmospheric turbulence aberrations with d/r₀≈3, the residuals of the system's wavefront root-mean-square calculated based on the SIR-Net model were reduced to 0.059λ, the inferring time is 0.191 ms, and the model size is 1.772 MB. Our proposed SIR-Net is expected to solve the problem of rapid wavefront detection in Hartmann under strong turbulence conditions.