Snapshot digital in-line holography with machine learning

Abstract

Digital in-line holography (DIH) has revolutionized applications in biomedical, physical sciences, and engineering by enabling the reconstruction of wavefronts from recorded interference patterns. DIH offers compactness and portability, making it a valuable imaging technique. However, the reconstruction process in DIH encounters obstacles, particularly with single-shot recording, which gives rise to twin-image effects and zero-order contamination. To surmount these challenges while preserving the simplicity and competitiveness of DIH, the development of advanced algorithms is imperative for snapshot reconstruction.

Reconstruction problems in holography can be classified as inverse problems, where the goal is to decode object information from encoded measurements. Recovering the target information from a single-shot in-line hologram is inherently ill-posed or underdetermined, presenting a significant challenge. Despite the progress made in snapshot DIH reconstruction, there is still ample room for further advancements, especially with the development of computational imaging techniques.

This dissertation focuses on addressing the reconstruction problems in snapshot DIH using advanced machine learning strategies. The research findings demonstrate the effectiveness and potential of advanced machine learning techniques in snapshot DIH reconstruction. The dissertation categorizes the methods into three groups: classic methods, deep methods, and physics-aware machine learning.

First, the superiority of one-stage networks over classic methods is demonstrated for real-time particle volumetric reconstruction from a single-shot hologram. This approach showcases the capability of deep learning techniques to outperform traditional methods in terms of speed and accuracy.

Next, the integration of physics information into neural networks through algorithm unrolling is explored. Specifically, the dissertation investigates DIH under photon-starved situations using quanta image sensors. By incorporating physics-based constraints into the deep learning framework, this approach enhances the interpretability and convergence guarantees of the reconstruction process.

Furthermore, the dissertation explores diffusion models for unsupervised DIH reconstruction, eliminating the need for paired holographic training datasets. This approach leverages the intrinsic properties of the holographic data to learn the underlying object information without explicit supervision, expanding the possibilities of unsupervised learning in DIH.

Lastly, the dissertation addresses the issue of model mismatch in holographic imaging under deterministic perturbations through a joint optimization framework. By jointly optimizing the hologram reconstruction algorithm and the physical model, this approach mitigates the impact of model mismatch and improves the accuracy of the reconstructed object

In conclusion, this dissertation highlights the effectiveness and potential of advanced machine learning techniques in snapshot DIH reconstruction. It emphasizes the importance of integrating physics information into deep learning frameworks and explores unsupervised learning approaches. The research aims to enhance the capabilities of snapshot DIH, promote its wider adoption, and contribute to various domains and industries. Future research directions are suggested to further advance snapshot DIH reconstruction and explore novel applications of this technology.