RS-N2N: a single-image phase denoising network for digital holographic microscopy

Phase noise significantly degrades the measurement accuracy of digital holographic microscopy (DHM). However, deep-learning-based denoising methods often require large-scale paired datasets that are difficult to obtain, and networks trained solely on Gaussian noise typically exhibit poor generalization, leading to unsatisfactory performance in removing practical noise. In this paper, we propose RS-N2N, a single-frame denoising network tailored for DHM phase maps. The network incorporates Fourier-based preprocessing and SE-block attention mechanisms and introduces a self-constrained learning strategy. With the aid of data augmentation, the model can be effectively trained using only a single noisy phase image. Experimental results demonstrate that RS-N2N outperforms previous methods by 1.3 dB in PSNR for Gaussian noise removal, achieves up to a 3 dB improvement for Perlin noise, and shows an average gain of approximately 1.5 dB in removing real phase noise. These results indicate that RS-N2N is highly effective for denoising phase noise in digital holographic microscopy.