Core Concepts
Proposing reverse-attention loss to address the Autofocusing challenge in untrained deep learning models for digital holography.
Abstract
The article discusses the challenges faced by untrained physics-based deep learning methods in digital holography due to uncertain object distances. It introduces reverse-attention loss as a solution to improve efficiency and accuracy in reconstructing holograms. The method is compared against conventional solutions and DL-based supervised methods, demonstrating superior performance. The theoretical analysis and experiments support the effectiveness of reverse-attention loss in addressing the Autofocusing issue.
The content is structured into sections covering Introduction, Background Information, Problem Formulation, Methodology, Synthetic Analysis of Convergence, Experiments on Samples, Conclusion, and Acknowledgments. Key insights include the explanation of DIH reconstruction challenges, the formulation of a continuous-discrete optimization problem, the proposal of reverse-attention loss, convergence analysis, experimental results showcasing improved reconstruction performance using the proposed method.
Stats
"For example, the difference is less than 1dB in PSNR and 0.002 in SSIM for the target sample in our experiment."
"We choose DeepDIH [19] as the baseline model in this work since DeepDIH is one of the most representative untrained physics-driven DL methods for DH reconstruction."
"Our method only takes 15 minutes."