Bahram Javidi

Unmanned Autonomous Vehicles (UAVs) are increasingly deployed across various applications, necessitating techniques for secure tracking and authentication against spoofing vulnerabilities. While physically unclonable function (PUF)-based optical ID tags offer secure signatures, conventional speckle acquisition systems struggle to extract full three-dimensional (3D) spatial coordinates of UAVs due to poor speckle sensitivity to longitudinal movements of UAVs. To overcome this problem, we propose a hybrid dual-path optical acquisition system by supplementing a security authentication system with a 3D integral imaging sensing system. This modification introduces two challenges: 1) data of mixed modality, which are both 2D and 3D videos, and 2) the possibility of the optical encryption key information leaking into the parameters of the processing pipeline. We propose a two-step framework where a 3D U-Net extracts 3D spatial priors before a separate spatio-temporal fusion network integrates these 3D spatial priors with the encrypted speckle signatures to infer temporal dynamics. The proposed separation of networks is designed to reduce the potential leakage of the optical key into the network weights by preventing the unencrypted speckle data from being jointly processed with the encrypted speckle patterns. Experimental validations performed in the lab demonstrate that our spatio-temporal integrated network maintains the 75 ^th percentile of the Euclidean tracking error below 1.1 voxels for precise 3D spatial localization and achieves a high F1-score across all test scenarios. The results confirm that the proposed framework has the potential to accomplish simultaneous 3D tracking and security authentication of UAVs, offering a spoofing-resistant solution based on physical optical encryption for advanced surveillance, defense, and autonomous navigation applications.

Digital holographic microscopy provides label-free imaging and quantitative phase information for microscopic particles in aquatic and fluidic environments. In this review paper, we review a number of digital holographic microscopy (DHM) configurations, including two-beam Mach-Zehnder, lateral shearing self-referencing, and lens-less systems for imaging, analysis, and classification of biological cells and microplastic particles. Phase-based features are used to classify morphologically similar yeast cells, achieving an accuracy of ∼90% using classical machine learning models. Lateral shearing DHM demonstrates improved temporal stability and enables three-dimensional tracking and imaging of diatoms and microspheres. A lens-less configuration is used for high-throughput imaging of highly scattering microplastics, where combined intensity and phase features enable classification of polystyrene and PMMA particles with ∼84% accuracy. The integration of digital holographic microscopy with microfluidics further enables real-time imaging and tracking under controlled flow conditions. The results show that reliable classification and characterization can be achieved using simple optical setups and engineered features, making DHM suitable for practical applications in aquatic and fluidic environments.

Recent progress in artificial intelligence (AI) and digital holographic microscopy (DHM) has enabled quantitative, label-free, and noninvasive cellular imaging with unprecedented precision. This review provides an overview of AI-driven DHM technologies that transform classical holographic phase reconstruction and cellular analysis into real-time, portable biomedical solutions. After outlining the optical and computational fundamentals of DHM and quantitative phase imaging, we describe how deep generative and diffusion models substantially enhance phase retrieval accuracy under noisy or single-shot conditions. We then summarize recent biomedical applications, integrating blood, cancer, and cardiac cell analyses into a unified framework of AI-assisted quantitative phenotyping. Deep and self-supervised learning approaches are shown to enable high-accuracy classification of red blood cells and cancer cells and label-free evaluation of cardiomyocyte contractility and drug response. The combination of AI-based reconstruction, self-supervised learning, and physics-informed modeling demonstrates robust performance even with limited labeled data. Finally, we discuss the system-level transition toward low-cost, edge-AI-enabled DHM platforms capable of real-time phase imaging in point-of-care or field environments. We highlight key challenges in data standardization, interpretability, and multimodal integration. Collectively, this review envisions AI-integrated DHM as a scalable, accessible technology bridging advanced quantitative imaging with practical biomedical diagnostics.

Digital holographic microscopy enables high-contrast, label-free quantitative phase imaging of transparent cells using visible light. This imaging modality can provide many biophysical and biomechanical cell features based on the numerically processed digital holograms. Sickle cell disease is a condition that affects red blood cells and is a significant health concern worldwide. Its clinical diagnosis, conventionally, is made using high-performance liquid chromatography, which is costly and time-consuming and requires experts to interpret the results. This paper presents a cost-effective method for screening sickle cell positive samples using a compact, self-referencing lateral shearing digital holographic microscope. Holographic videos of healthy and sickle cell positive samples are recorded using a digital holographic microscope and are numerically reconstructed. Both spatial and temporal features of the cells are extracted from the recorded holographic videos. The distribution of these features indicates the difference between healthy and sickle-cell-positive red blood cells. These features are used with classical machine learning classifiers to evaluate classification accuracy. The results are compared to determine the best machine learning model for label-free sickle cell identification. Results show that simple linear classifiers are sufficient due to the physical separability of the extracted features.

Astrocytes are star-shaped glial cells whose elaborate soma-process architecture supports neuronal functions, regulates ions as well as water homeostasis, and contributes to brain structural organization. This paper presents a quantitative imaging technique for comprehensive morphological analysis of astrocytes, including their somata and branches, using a custom-built common-path quantitative phase microscope. The microscope employs a common-path off-axis configuration, splitting the sample beam using a grating, and a dual-lens system combined with a spatial filter to produce a reference beam required for hologram generation. This microscope enables a label-free, high-temporal-stability, and real-time 3D visualization of biological samples by measuring optical path length difference (OPD). Astrocyte cells were imaged, and their morphometric properties were quantified in near real time using the developed microscope. The reconstructed optical path length difference (OPD) maps were used to perform multistage deep learning segmentation of entire astrocyte cells, and the corresponding soma and branches were segmented using an optical thickness-based clustering approach. The segmented soma and process regions were compared as well as correlated by their phase-derived and geometrical morphological parameters. To the best of our knowledge, this is the first report of using the proposed digital holographic approach with an AI model to segment and measure the morphometrics of astrocyte cells, including their somata and branches separately.

We evaluate a three-dimensional (3D) object detection system for operation in turbid water based on 3D profilometric Integral Imaging (InIm) and a deep neural network. While conventional InIm computational reconstruction provides the two-dimensional (2D) slices of the 3D scene at specific 2D depth planes, 3D profilometry allows visualization of the 3D surface from unique perspectives. In the proposed method, we develop a deep neural network-based red, green, blue-depth (RGB-D) object detection framework using passive 3D profilometry under turbid conditions. An image sensor on a moving platform captures multiple 2D perspective images of the 3D scene, from which a depth map is statistically estimated. The captured perspective image and the estimated depth map are then fused to generate a four-channel RGB-D image for 3D object detection in turbidity. Comparative experiments are conducted and demonstrate that the proposed 3D profilometry-based approach outperforms both 2D imaging and conventional 3D InIm-based reconstruction across various turbidity levels evaluated. To the best of our knowledge, this is the first report on InIm 3D profilometry for object detection in turbid water.

There are many applications for secure authentication of unmanned moving vehicles (UMVs). Traditional object appearance-based detection systems might be vulnerable to adversarial optical deception (spoofing). To address this challenge, we crafted a tracking and authentication system resistant to spoofing by equipping both the unmanned moving vehicles and the optical receiver with optical ID tags. A variety of approaches may be used to generate secure optical tags, such as three-dimensional (3D) phase encoded tags or tags utilizing meta materials. The optical tag mounted on the UMVs generates a unique optical signature produced by illuminating the ID tag with a light source. At the receiver, this optical signature is optically encrypted by a proprietary optical key placed within the optical acquisition system to ensure that the signature of the authorized UMVs cannot be reverse engineered. The encrypted signature is then processed by a deep learning network, such as the UNet-ConvLSTM network, within a conditional generative adversarial network (cGAN) framework. This network is trained to authenticate the encrypted unique signature of the tag and track its location simultaneously. We experimentally evaluate the performance of the present method under three environmental conditions, such as normal light, low light, and fog. The results show that, through training for each environment, our model successfully tracks and authenticates the optical ID tags even under degraded environmental conditions. Our system might be a solution for spoof-resistant tracking and authentication of UMVs such as unmanned aerial vehicles, drones, aircraft, satellites, etc. in various applications, including defense and surveillance. To the best of our knowledge, this is the first report on secure authentication and tracking of unmanned moving vehicles with optical ID tags.

Underwater optical signal detection in severely degraded environments remains a challenging task. In this work, we propose what we believe to be a novel dynamic vision-based underwater optical signal detection system. The signal detection performance is evaluated in a degraded underwater environment, including partial occlusion and turbidity. The system utilizes dynamic vision sensors from an event camera array, multi-dimensional integral imaging, and deep learning networks to achieve signal detections. In the experiment, optical signals are transmitted using a modulated light-emitting diode. The optical signals, after propagating through the degraded underwater environment, are captured by the event camera array in the form of event sequences. The event sequences are preprocessed as multi-dimensional event videos. The videos are classified by the vision transformer and gated recurrent unit network (ViT-GRU). The proposed system is compared to other relevant state-of-the-art frame-based approaches in terms of the detection performance evaluated by the Matthew correlation coefficient and the number of error symbols. For the experiments we conducted, the proposed dynamic vision-based underwater optical signal detection system with multi-dimensional integral imaging and ViT-GRU network outperforms other frame-based counterparts in degraded underwater environments. To the best of our knowledge, this is the first report on dynamic vision-based underwater optical signal detection using multidimensional integral imaging.

We propose an underwater multidimensional metrology approach in degraded environments utilizing augmented reality devices. The approach uses a risk-aware next-best-view (NBV) framework for metrology in degraded visual environments, introducing a conditional value-at-risk (CVaR) objective to guide viewpoint selection. Using anisotropy-aware geodesic metrics, local sensitivities, and a CVaR optimization, the approach prioritizes views that reduce worst-case geodesic-length uncertainty. We validated the performance of the proposed system experimentally on turbid scenes with augmented reality device imagery and five-dimensional WaterSplat reconstructions. The results across increasing turbidity demonstrate that CVaR achieves 2 − 3 × greater median error reduction compared to entropy-based NBV selection. We further conduct controlled synthetic experiments to analyze asymptotic error plateaus. To the best of our knowledge, this is the first report on an augmented reality-based NBV approach aimed at improving metrology in turbid environments.