Physics-Informed Decoupled Calibration for Fourier Ptychographic Microscopy

Fourier ptychographic microscopy (FPM) is a promising quantitative phase imaging technique with large fields of view and high resolution, but it requires precise illumination angles for accurate reconstruction. Conventional algorithms struggle to rapidly separate system errors and impose strict constraints on imaging systems. To address this, we propose a physically decoupled correction framework integrating convolutional neural network (CNN), simulated annealing (SA) algorithms, and GPU parallel acceleration. The CNN extracts frequency-domain circular features related to LED positioning errors as physical priors, while the GPU-accelerated SA algorithm accurately solves LED array spatial parameters during FPM forward propagation. Because this method is decoupled from phase recovery, single-round calibration parameters apply to diverse conditions, reducing error correction time by >67.7% and improving imaging efficiency by >60.1%. Experiments verify its ability to precisely calibrate LED positions, enhancing FPM robustness and laying a solid algorithmic foundation for efficient full-field error correction.