SNR enhancement for low-SNR amplitude-modulated holographic data storage based on deep learning

In amplitude-modulated holographic storage, due to material inhomogeneity, optical system aberrations, and environmental interference, the recorded and read amplitude data pages often contain a large amount of random noise, leading to a decrease in the image signal-to-noise ratio (SNR). However, the traditional U-Net has a limited ability to handle random noise, particularly for low-SNR data pages, making it difficult to effectively enhance SNR. This paper proposes an improved U-Net named DRAMCU-Net (dilated residual attention and multi-scale convolution U-Net) for enhancing the SNR of holographic data storage images. DRAMCU-Net achieves multi-level feature extraction and attention focusing by introducing dilated residual attention blocks (DRABs) and utilizing multi-scale convolutional blocks (MS-Conv) instead of traditional convolutional blocks to efficiently capture feature information at different scales. Additionally, a dropout layer is employed to enhance the model's global robustness. Experimental results demonstrate that compared to the U-Net, the low-SNR data pages reconstructed by DRAMCU-Net achieve approximately 30% higher SNR.