High-quality computer-generated holography based on Vision Mamba

Abstract

Deep learning, especially through model-driven unsupervised networks, offers a novel approach for efficient computer-generated hologram (CGH) generation. However, current model-driven CGH generation models are primarily built on the convolutional neural networks (CNNs), which struggle to achieve high-quality hologram reconstruction due to limited receptive fields. Although Vision Transformers (ViTs) excel at processing more distant visual information, they are burdened with huge computational load. The recent emergence of Vision Mamba (ViM) presents a promising avenue to address these challenges. In this study, we introduce the CVMNet, a lightweight model that combines the precision of convolutional layers for local feature extraction and the long-range modeling abilities of state-space models (SSMs) to enhance the quality of CGHs. By employing parallel computation for the ViM to handle feature channels, the CVMNet effectively reduces the number of model parameters. Numerical reconstruction and optical experiments demonstrate that the CVMNet can generate 1080P high-quality holograms in just 16 ms, boosting an average PSNR of over 30 dB and effectively suppressing speckle noise in reconstructed images. Additionally, the CVMNet showcases robust generalization capabilities.