Deep Learning-based Methods for Low-noise Hologram Generation
SignificanceHolographic 3D display can reconstruct the complete wavefront information of 3D objects and is therefore regarded as a key direction for the development of next-generation naked-eye 3D display. However, traditional hologram generation methods struggle to balance holographic reconstruction quality with calculation speed. Among these challenges, suppressing speckle noise is a core challenge in holographic 3D display. Originating from coherent illumination and random phase distributions, speckle noise is the primary factor affecting the quality of holographic reconstructions. In recent years, deep learning has provided a new computational paradigm for generating low-noise holograms. Therefore, this paper systematically reviews deep learning-based methods for generating low-noise holograms, providing a detailed discussion of the basic principles, technical advancements, and future trends of data-supervised, model-supervised, and hybrid-supervised methods, with the aim of providing a reference for further optimizing low-noise hologram generation methods.ProgressIn recent years, low-noise hologram generation methods based on deep learning are experiencing rapid development, and innovations around data-supervised, model-supervised, and hybrid-supervised methods are driving the progress of holographic 3D display. Data-supervised methods are dependent on paired datasets of target images and holograms, and the low-noise hologram is generated by learning the mapping relationship between them. Early data-supervised studies are focused on the application of deep neural networks to hologram generation, where tensor holography enables real-time synthesis of low-noise holograms, and diffraction-engineered networks further enhance the visual realism of holographic 3D display. A noteworthy strategy is that low-noise holograms are generated directly from 2D target images, which simplifies the construction of training datasets. However, data-supervised methods remain highly dependent on large-scale, high-quality datasets, and both generalization ability and model performance are closely related to dataset quality. Model-supervised methods embed differentiable physical models into the training process, which not only avoids reliance on paired datasets but also enables unsupervised or self-supervised learning of low-noise holograms by minimizing the intensity difference between reconstructed and target images. Hardware-based model-supervised methods calibrate optical aberrations through system feedback and achieve high-quality color holograms. Software-based model-supervised methods introduce complex-valued neural networks or lightweight architectures to generate holograms with high accuracy and high frame rates. In addition, model-supervised methods are applied to solve occlusion problems in complex 3D scenes. Nevertheless, repeated diffraction propagation calculations are required in the model-supervised methods, and the training cost is high. Hybrid-supervised methods integrate the advantages of data and model supervision, and generate high-fidelity, low-noise holograms by combining data and physical losses. For example, two-stage hybrid training method first pre-trains networks on high-quality datasets and then performs unsupervised fine-tuning with physical models, thereby surpassing the performance ceiling of data-supervised methods. Under data-scarce conditions, hybrid-supervised methods based on non-paired learning achieve object-to-hologram mapping through cycle-consistency structures. Hybrid-supervised methods are also applied to optical system optimization, propagation error compensation, and hologram generation under low spatial coherence. However, hybrid-supervised methods still face challenges of complex training processes and limited stability.Conclusions and ProspectsAlthough deep learning drives the rapid development of holographic 3D display, several challenges are still encountered: 1) in data-supervised methods, the training efficiency of deep neural networks is low and strong dependence on training datasets is observed; 2) in model-supervised methods, the speed of hologram generation does not yet meet the requirements of real-time interaction; 3) in hybrid-supervised methods, the gain obtained by simply increasing the depth or width of neural networks becomes limited. To address these issues, future research is expected to proceed along the following directions: 1) transfer learning methods are applied to reduce the training time of data-supervised methods and to alleviate dependence on datasets; 2) lightweight neural networks are designed to improve the generation speed of model-supervised methods; 3) novel architectures such as Transformer models, self-attention mechanisms, and diffusion models are explored to extend the performance boundaries of networks. With improvements in training efficiency, inference speed, and reconstruction quality, holographic 3D display is expected to promote transformative applications in medical imaging, industrial inspection, education, and entertainment.