Faster Dictionary Learning Method Speeds Up Multi-Beam Holographic Microscopy

Digital holographic microscopy has long offered biologists and engineers a way to image transparent, delicate specimens in three dimensions without stains or dyes, but the technique carries a heavy computational price. Reconstructing a hologram requires numerical wave propagation, phase retrieval, and full three-dimensional reconstruction, all of which depend on large-scale matrix computations. As spatial resolution and sampling density rise, that burden grows substantially, often forcing researchers to choose between image fidelity and reconstruction speed. A new study published in Applied Physics B proposes a way to ease this trade-off by integrating an efficient dictionary learning algorithm, called Sequential Generalized K-means (SGK), into the reconstruction pipeline of a multi-beam holographic microscope. The authors report that the approach achieves high-fidelity reconstructions and faster convergence than conventional dictionary-learning approaches, making it a candidate for real-time microscopic imaging applications.

The work builds on the Light-Origami Multi-Beam Interference Digital Holographic Microscope, or LOMBI-DHM, a system recently introduced to address one of holography’s persistent constraints: the field of view is limited by the size of the image sensor. LOMBI-DHM doubles the field of view by folding two spatially distinct object regions onto a single sensor using a cube beam splitter. The interference of two object beams and a common reference beam yields a multiplexed hologram that can be numerically reconstructed to recover amplitude and phase from both fields of view simultaneously. According to the authors, unlike earlier prism-based or reflective-path systems, this dual-channel transmission interferometer offers high light throughput, reduced power requirements, and a doubled field of view in a single exposure, at the cost of lower temporal phase stability.

The optical layout described in the paper is intricate but purposeful. A 632.8 nm helium-neon laser is expanded, spatially filtered through a 40x microscope objective and a 5 micrometre pinhole, and collimated with a 100 mm focal length lens. A 50:50 cube beam splitter divides the collimated beam into object and reference arms. The object, mounted on a three-dimensional translation stage, is illuminated by the object beam, and its diffracted light is collected by a microscope objective and recollimated with another 100 mm lens. A second beam splitter, oriented at 45 degrees on a three-dimensional stage, then generates two identical but inverted fields of view carrying distinct object information. If half the field contains the letter A and the other half the letter B, the splitter routes them into separate triangular prisms, producing two overlapped yet inverted fields at its output. These are combined with the reference beam at a third, slightly tilted beam splitter to realize off-axis interference, forming a multiplexed hologram recorded by a Sony Pregius IMX264 sensor with 2448 by 2048 pixels and 3.45 micrometre pixel size.

The researchers first validated the system with quantitative phase imaging experiments on a USAF 1951 resolution target, imaged with a 20x microscope objective of numerical aperture 0.40. Two distinct regions of the target, elements 3 and 4, were simultaneously encoded within the sensor’s field of view. Because both regions share a common reference beam, each object beam pair contributes a unique complex wavefront. In the spatial frequency domain, the multiplexed hologram reveals three off-axis interference terms: two from the interaction of the reference beam with each field of view, and a third arising from interference between the two fields themselves. From a single hologram, the complex amplitude distributions of both fields of view could be retrieved numerically, with reconstructed amplitude images and corresponding phase maps confirming the dual-channel recovery.

The core contribution, however, lies in the reconstruction mathematics. Dictionary learning rests on the observation that sparsity is a fundamental property of many natural signals, and exploiting it can achieve dimensionality reduction, lowering storage and computational requirements. Unlike compressive sensing, which primarily addresses efficient signal acquisition and recovery, dictionary learning constructs adaptive dictionaries that represent the underlying data more suitably than predefined transform bases such as Fourier or wavelets. In the proposed framework, the input hologram is divided into overlapping patches, and each patch is expressed as a linear combination of sparse dictionary atoms drawn from an overcomplete dictionary. The dictionary and sparse codes are jointly estimated by minimizing the reconstruction error subject to a prescribed sparsity constraint that limits the number of nonzero coefficients in each patch’s representation.

The distinguishing feature of SGK is how the dictionary is updated. The learning process is initialized using a Fast Discrete Curvelet Transform (FDCT) basis, and sparse coding is performed with Orthogonal Matching Pursuit. For each dictionary atom, the algorithm identifies the support set of training patches that use that atom, then updates the atom by aggregating, through simple arithmetic averaging, the patches associated with it. The updated atom is normalized to unit Euclidean norm, preventing scaling ambiguity between atoms and their coefficients. This stands in sharp contrast to the widely used K-SVD algorithm, whose alternating sparse coding and dictionary update steps rely on repeated singular value decompositions. Those decompositions incur substantial computational overhead, limiting scalability to large holographic datasets, and SVD-based iterative methods may exhibit instability and excessive runtime when reconstruction is sensitive to noise characteristics and sparsity constraints. By replacing SVD with averaging, SGK significantly reduces algorithmic complexity while preserving adaptive dictionary learning capability.

To quantify performance, the authors evaluated reconstructions using standard image-quality metrics. Mean squared error measures the average squared difference between the reference holographic image and the reconstruction, with lower values indicating higher accuracy. Peak signal-to-noise ratio compares the maximum possible signal intensity with the reconstruction error, where higher values indicate better quality and lower distortion. The structural similarity index measure evaluates perceptual similarity by considering luminance, contrast, and structural information, ranging from minus one to one, with values closer to one indicating higher structural similarity. Execution time, number of iterations, and memory consumption were also reported for each method to provide a fair assessment of computational efficiency. All simulations were performed in MATLAB R2020b on a system equipped with an Intel Xeon Silver 4216 CPU at 2.10 GHz, with two processors, and 256 GB of RAM.

The experimental tuning revealed the importance of parameter selection. The team tested various patch and dictionary sizes and found that a patch size of 3 by 9 provided the best performance, a configuration then adopted for all compared methods. Three sparsity levels were considered, T equal to 1, 2, and 3. For a 200-atom dictionary, these levels correspond to approximately 0.5, 1.0, and 1.5 percent active nonzero coefficients, while the remaining 99.5, 99.0, and 98.5 percent of coefficients are constrained to zero. As the sparsity level increased from 1 to 2 and then to 3, reconstruction quality improved noticeably, producing clearer and sharper intensity patterns, which indicates that a higher sparsity level allows the dictionary to represent the holographic data more accurately and reduces reconstruction error. The proposed method was compared against K-SVD, OMP-only reconstruction, and FDCT dictionary reconstruction under identical experimental conditions, with the FDCT-initialized SGK framework reported to achieve superior reconstruction quality while significantly reducing computational complexity.

A practical advantage emphasized by the authors is that the method preserves the standard off-axis DHM acquisition configuration and improves only the numerical reconstruction stage. Conventional off-axis reconstruction generally proceeds through Fourier-domain spectrum filtering followed by numerical wave propagation. In contrast, the SGK-based strategy leverages sparse representation of holographic data to obtain a more reliable reconstruction, with the adaptive dictionary optimized to represent the intrinsic spatial features of the hologram, suppressing noise and reconstruction artifacts while retaining subtle intensity and phase information. Because the approach operates entirely in the computational reconstruction stage, it can be incorporated into existing off-axis DHM systems without any changes to the optical recording arrangement. The patch-based aggregation also ensures consistency across overlapping regions, improving reconstruction stability and reducing boundary artifacts.

The authors are careful to frame the work as promising rather than finished. Future work will focus on optimizing the dictionary learning process to reduce computational complexity and reconstruction time further, aiming toward real-time or near-real-time holographic reconstruction, and on integrating adaptive and online dictionary learning techniques to handle diverse object types and varying imaging conditions. Robustness under challenging environments, particularly low illumination and significant noise contamination, remains an open challenge: limited photon availability, detector-induced noise, and speckle interference often deteriorate hologram quality and reduce the accuracy of quantitative phase recovery. Planned directions include noise-adaptive dictionary learning frameworks, physics-guided deep learning strategies, improved denoising algorithms, and more effective regularization schemes for photon-starved acquisition settings, conditions frequently encountered in biological microscopy where reduced illumination is needed to minimize phototoxicity and photobleaching during prolonged imaging. If those efforts succeed, the researchers anticipate improvements that could support quantitative phase imaging systems for biomedical applications, including more reliable analysis, disease diagnosis, and long-term observation of living specimens.

Subject of Research: A sparse dictionary learning algorithm for faster reconstruction in multi-beam digital holographic microscopy

Article Title: Dictionary Learning-Based Multi-Beam Digital Holographic Microscopy

Article References: Lakshmi, K., Dodda, V. C., Kumar, M., Kumar, R., Muniraj, I., & Yeo, I. (2026). Dictionary Learning-Based Multi-Beam Digital Holographic Microscopy. Applied Physics B, 132(9), Article 98. https://doi.org/10.1007/s00340-026-08708-y

Image Credits: AI Generated

DOI: 10.1007/s00340-026-08708-y

Keywords: digital holographic microscopy, dictionary learning, Sequential Generalized K-means, sparse reconstruction, K-SVD, phase imaging, LOMBI-DHM, multiplexed hologram, computational imaging, biomedical imaging, Orthogonal Matching Pursuit, reconstruction algorithms