Coherent Noise Suppression of Single-Shot Digital Holographic Phase Via an Untrained Self-Supervised Network

Abstract

In digital holography, the coherent noise affects the measurement accuracy and reliability greatly due to the high spatial and temporal coherence of the laser. Especially, compared with the speckle noise of intensity in digital holography, the coherent noise of phase contains more medium- and low-frequency characteristics, which hinders the effectiveness of noise suppression algorithms. Here, we propose a single-shot untrained self-supervised network (SUSNet) for the coherent noise suppression of phase, requiring only one noisy phase map to complete the optimization and learning. The SUSNet can smoothen and suppress the background fluctuations, parasitic fringes, and diffraction loops in a noisy phase and shows good generalization performance for samples with different shapes, sizes, and phase ranges. Compared with the traditional algorithms and the ground truth-supervised neural network (DnCNN), the SUSNet has the best noise suppression performance and background smoothing effect. As a result, the SUSNet can suppress the fluctuation range to ∼20% of the original range.

On the use of deep learning for phase recovery

“…: 4000 pairs GAN loss Murdaca et al 217 Noisy real, imaginary, and amplitude Noise-free real, imaginary, and amplitude U-Net Sim. : 5400 pairs l 2 -norm Tang et al 219 Fixed noise matrix Noise-free phase U-Net (untrained) Expt. : 1 l 2 -norm, gradient, and variance Resolution enhancement Liu et al 220 LR phase and amplitude HR phase and amplitude U-Net Expt.…”

“…The difference is that in addition to the sine and cosine images of the phase, the neural network also reduces noise for the amplitude images at the same time. Tang et al 219 proposed to iteratively reduce the coherent noise in phase with an untrained U-Net. In the above works, various loss functions were employed alongside the conventional l 2 -norm and l 1 -norm to enhance performance.…”

“…In the above works, various loss functions were employed alongside the conventional l 2 -norm and l 1 -norm to enhance performance. These additional functions include the edge function 208 , which sharpens the edges of the denoised image, as well as gradient and variance functions 219 that further suppress noise while preventing excessive smoothing.…”