Rapid computer-generated hologram with lightweight local and global self-attention network

Abstract

Computer-generated holography is a technique that utilizes computers and algorithms to generate holographic images. Deep learning-based Computer-generated holography can learn the mapping relationship between input images and holographic images, which offers faster computation speed and better image quality comparing with direct encoding and iterative optimization methods. However, most methods typically employ stacked convolutional layers to expand the receptive field, which leads to a sharp increase in computational cost as well as number of parameters. We proposed a rapid computer-generated holograms method with lightweight local and global self-attention networks (LGSANet), which performs phase encoding of input images as an alternative to the conventional holographic method for recording object information. Once the network training is completed, it is possible to perform high quality holograms with a spatial resolution of 1920 × 1080 within 39 ms. By importing the encoded phase of input images into a spatial light modulator (SLM), a clear reconstructed image can be obtained at the observation plane by irradiating the SLM with a reference light (@λ = 532 nm). Experimental results show the proposed method exhibits significant improvements in peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), as well as reduced speckle noise, which can be applied to holographic displays, AR/VR, metasurface design, medical imaging, etc.