A phase retrieval algorithm combines deep learning for testing optical surface

This paper presents a hybrid phase retrieval algorithm that combines deep learning and traditional algorithms. During the process of detecting large dynamic range, the problem of phase wrapping arises, which cannot be solved by traditional iterative algorithms. Firstly, Zernike polynomials are used to represent the low-frequency component of mirror errors, and a CNN neural network model is trained to predict the coefficients of Zernike polynomials. Then, the high-frequency component of the error is retrieved using the GS algorithm. In the simulation, the tested spherical mirror is set with PV 3.87λ, RMS 0.267λ. The recovered phase wavefront error is 4.098λ, with RMS 0.258λ, and the RMS relative error is 3.21%.