Noise Suppression and Unpacking Accuracy Improvement Based on Generative Adversarial Networks for Interferometric Phase Maps

Abstract

Multiplicative speckle noise and additive Gaussian noise in the interferometric phase map introduce spurious phase jumps and residual points, disrupting phase continuity and reducing the reliability of the phase unpacking algorithm. To address this issue, this paper proposes a conditional generative adversarial network. The generator uses a multi-scale dilated convolutional U-Net to output a complex phase residual map. The denoised result is obtained by subtracting the residual from the noisy phase. A dual discriminator structure in the spatial and frequency domains, combined with weighted structural similarity loss, phase gradient consistency loss, and least-squares adversarial loss, jointly constrains the network training. Training data is generated through electromagnetic scattering simulation. The proposed method achieves a 97.3% phase unpacking success rate at a signal-to-noise ratio of 0–3 dB. Processing a 2048 × 2048 pixel image takes less than 0.25 seconds. The proposed method effectively suppresses non-stationary noise and improves the accuracy of phase unpacking.