Research on Acceleration Algorithms for Computer-Generated Holography Based on Multimodal Deep Learning

Three-dimensional holographic imaging serves as a core enabling technology for cutting-edge domains including the metaverse, immersive virtual reality, precision medical imaging, industrial numerical simulation, and information visualization. The computational efficiency of Computer-Generated Holography (CGH) directly determines the real-time performance and practical viability of 3D holographic imaging systems. CGH relies on large-scale numerical simulation of optical interference and diffraction, characterized by ultra-high computational complexity and intensive computing demands. Conventional serial computing architectures fail to satisfy the requirements of high-resolution and large-scale applications. In recent years, continuous advances in multimodal deep learning, distributed cloud computing, and GPU heterogeneous parallel computing have opened up a new technical route for CGH acceleration. In this study, a distributed parallel computing framework for CGH is constructed based on Baidu PaddlePaddle deep learning cloud platform. On this basis, physics-constrained multimodal feature fusion, adaptive dynamic task scheduling, and three cutting-edge technologies are innovatively integrated: Mamba state-space model for high-speed hologram inference, 4D hybrid parallel distributed computing, and edge–cloud collaborative heterogeneous acceleration. The acceleration performance, imaging quality stability, and adaptive scheduling efficiency of the proposed method are systematically tested and evaluated. Experimental results demonstrate that, compared with the standalone serial computing mode based on the traditional Pentium 4 processor, the proposed method delivers a maximum speedup of 1426 times, while improving the average peak signal-to-noise ratio (PSNR) of reconstructed images by 3.2 dB. The adaptive scheduling scheme reduces communication overhead by 18.7% under fluctuating computing loads, effectively breaking through the long-standing trade-off bottleneck between CGH computational speed and imaging quality. This work provides an efficient, feasible technical solution for the real-time, engineering, and large-scale deployment of 3D holographic imaging.