Advances in Hologram Reconstruction: A Survey of Traditional and Deep Learning Techniques

Hologram reconstruction retrieves 3D information from recorded holograms for applications in biomedical imaging, augmented reality, industrial inspection, and security. Traditional methods, including Fourier transform-based techniques and iterative phase retrieval, face challenges such as noise sensitivity and high computational costs. In contrast, deep learning approaches—such as CNNs, GANs, and transformers—have significantly improved reconstruction accuracy, efficiency, and real-time applicability. This survey comprehensively reviews 66 recent studies, categorizing methods into traditional and deep learning paradigms, and systematically analyzes their methodologies, datasets, performance, and limitations. By incorporating emerging transformer architectures, self-supervised learning strategies, and hybrid AI-physics models, this work provides a holistic view of current trends and research gaps. Key observations indicate that transformer-based and hybrid approaches offer superior noise robustness and scalability, paving the way for real-time, high-fidelity holographic imaging. Future research directions include model optimization, hardware acceleration, and cross-domain applications to further enhance computational efficiency and reconstruction quality.