Phase unwrapping in digital holographic interferometry via physics-informed convolutional-Fourier neural network

Phase unwrapping is essential for many phase-centric optical imaging and metrology techniques, yet remains a challenge under high speckle noise. A deep learning-based speckle phase unwrapping algorithm was developed on the basis of a physics-informed convolutional-Fourier neural network, relying on an uncertainty-aware mathematical model of speckle. To address the limitation that traditional spatial-domain CNN methods suffer from inadequacy in extracting frequency-domain interference features during phase unwrapping under speckle noise, we integrated the speckle mathematical model into the network via a parallel convolution-Fourier architecture and employed an integrated loss function with speckle noise, phase gradient, and absolute phase to significantly enhance the unwrapping performance of the network. Experimental results demonstrate that the proposed method achieves superior performance over other phase unwrapping algorithms in both simulations and digital holography experiments. By unifying spatial-domain and frequency-domain feature extraction via a physics-informed deep learning architecture, this method exhibits great potential for extension to broader information processing tasks.