Computational holographic imaging enables three-dimensional reconstruction by recording the amplitude as well as frequency phases of light waves. However, reconstructed images often suffer from coherent noise, speckle artifacts, and loss of fine structural details, limiting their practical applications in microscopy, biomedical imaging, and 3D visualization. This research proposes a deep learning-based approach for noise suppression and detail enhancement in computational holographic image reconstruction. The MNIST with 70,000 images and Text data was preprocessed using a Gaussian filter and data augmentation. A Queuing Search-driven Denoise Adaptive Residual Dense Network (QS-Denoise ARDN) model is trained directly on distorted phase images, eliminating the need for uninterrupted ground truth data, while a noise level function network estimates local noise characteristics. The networks are jointly optimized by maximizing a constrained negative log-likelihood function, enabling effective suppression of coherent and speckle noise. Experimental evaluation on various holographic datasets demonstrates that the proposed method significantly improves image quality compared to conventional smoothing and phase recovery algorithms. Quantitative analysis shows marked improvements in the MNIST dataset (PSNR 26.84 dB) and SSIM (0.97). Text data (PSNR 27.48 dB) and SSIM (0.96), confirming the efficacy of the approach with the simulation by Python 3.10. Moreover, the model achieves rapid reconstruction with fewer measurements, highlighting its potential for real-time applications. These results indicate that integrating self-supervised deep learning with neural-network-based holographic reconstruction provides a robust and efficient solution for high-fidelity computational holography, offering new avenues for advanced biomedical imaging, optical metrology, and 3D visualization systems.
Fangju Li (2026) studied this question.