A high-fidelity Fresnel phase hologram generation method based on the neural network

Abstract Computer-generated holography (CGH), which enables dynamic generation of arbitrary complex threedimensional wavefronts, has emerged as a core technology for near-eye 3D displays and holographic micronano fabrication. Current deep learning-based CGH methods mitigate the inherent time-quality tradeoff of conventional iterative algorithms, yet existing physics-model-driven networks still suffer from high-frequency information loss and limited receptive fields. To address these drawbacks, this work proposes a jointoptimization complex-valued high-frequency attention pyramid pooling network (CHAPP-Net) for Fresnel hologram generation. The proposed network incorporates an atrous spatial pyramid pooling (ASPP) module and a high-frequency feature attention mechanism (HF-AM), which substantially expands the network receptive field and augments high-frequency details of reconstructed holographic images. Combined with physical forward propagation modeling, fidelity loss terms, smooth regularization, and multi-prior physical constraints for image uniformity, the proposed receptive-field-enhanced network realizes joint optimization of phase holograms for incident field modulation. This optimization strategy alleviates feature-specific overfitting and improves network generalization performance. Both numerical simulations and optical experiments validate the superiority of CHAPP-Net in edge detail preservation, speckle noise suppression and reconstructed image uniformity optimization. Evaluated on the DIV2K validation set, CHAPP-Net achieves a peak signal-to-noise ratio (PSNR) of 37.62 dB and a structural similarity index measure (SSIM) of 0.958. The proposed method provides a feasible and promising solution for practical holographic display and holographic nanofabrication applications.