Two-dimensional phase unwrapping and denoising network based on frequency-domain operators and sparse attention

Two-dimensional phase unwrapping is critical in optical interferometry, digital holography, and synthetic aperture radar. However, traditional methods are prone to error propagation and excessive smoothing under complex conditions like noise and phase discontinuities. To address this, we propose FSA-PU, a lightweight deep learning model based on U-Net. It integrates multi-scale context modeling, sparse global dependency, and frequency-domain reconstruction to enhance accuracy and robustness. With depthwise separable convolutions, it further improves edge and detail recovery. Experiments on synthetic and real datasets show that FSA-PU achieves excellent performance in terms of PSNR, RMSE, and SSIM, even in challenging conditions.