Computer-generated holography based on DPAC and residual attention network

Addressing the challenges of speckle noise suppression and high-fidelity reconstruction in computed holography (CGH), this paper presents a novel DPAC Residual Attention Network (DRANet) framework that integrates dual-phase amplitude coding (DPAC) with a residual attention network. By leveraging physics-driven and data-cooperative optimization, the proposed method achieves efficient and high-quality hologram generation. Specifically, DRANet combines the physical encoding advantages of DPAC with the feature enhancement and noise suppression capabilities of the residual attention network. Initially, DPAC is employed to generate an initial phase hologram, thereby retaining its computational efficiency and physical interpretability. Subsequently, the residual attention network refines the hologram. Experimental results from both numerical simulations and optical reconstructions demonstrate that DRANet significantly reduces speckle noise and enhances the reconstruction quality of holograms, providing a promising solution for real-time, high-fidelity holographic displays that effectively balances physical rigor with the generalization capability of deep learning.