This PDF file contains the front matter associated with SPIE Proceedings Volume 13136, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

The ongoing paradigm shift in healthcare towards personalized and precision medicine is posing a critical need for noninvasive imaging technology that can provide quantitative tissue and molecular information. Magnetic resonance signals from biological systems contain information from multiple molecules and multiple physical/biological processes (e.g., T1 relaxation, T2 relation, diffusion, perfusion, etc.). So, magnetic resonance imaging (MRI) is inherently a high-dimensional imaging technology that can acquire structural, functional and molecular information simultaneously. In practice, due to the curse of dimensionality, MRI experiments are often done in a low-dimensional setting to acquire biomarkers one at a time. Such a “divide-and-conquer” approach not only reduces data acquisition efficiency but also makes it difficult to obtain molecular information in high resolution. By synergistically integrating machine learning with sparse sampling, constrained image reconstruction and quantum simulation, we have successfully demonstrated ultrafast high-dimensional imaging of the brain. This talk will give an overview of this unprecedented omni imaging technology and show some exciting experimental results of brain function and diseases.

Recurrent nuisance flooding is common across many parts of the globe and causes extensive challenges for drivers on the roadways. The prevailing monitoring methods for roadway flooding are costly and not automated or effective. The ubiquity of visual data from cameras and advancements in computing such as deep learning may offer cost-effective methods for automated flood depth estimation on roadways based on reference objects such as cars. However, flood depth estimation faces challenges due to the limited amount of data annotated with water levels and diverse scenes showing reference objects at various scales and perspectives. This study proposes a novel deep learning approach to automated flood depth estimation on roadways. Our proposed pipeline addresses variations in object perspective and scale. We have developed an innovative approach to generate and annotate flood images by manipulating existing image datasets of cars in various orientations and scales to simulate four floodwater levels for augmenting real flood images. Furthermore, we propose object scale normalization for our reference objects (cars) to improve water level predictions. The proposed model achieves an accuracy of 74.85% and F1 score of 74.32% for four water levels when tested with real flood data. The proposed approach substantially reduces the time and labor required for labeling datasets while addressing challenges in perspective/scale, offering a promising solution for image-based flood depth estimation.

The individual identification of sea turtles through manual and semi-automated methods, such as those used in current photo-identification algorithms like WildID, Hotspotter, and I3S-Pattern, requires images of acceptable quality. This process involves analyzing the geometric morphology of the facial and dorsal scales, where users must manually select the region of interest. However, this methodology faces challenges due to the inherent natural variability of these species and the morphological changes that occur over time. In image processing, automatic semantic segmentation techniques accurately delineate specific areas within images. This enables the detection, identification, and differentiation of objects through pixel-level analysis and visual pattern recognition, highlighting contours and specific object characteristics. In this work, we propose an artificial intelligence model and deep learning approach designed for the individual identification of sea turtles using multiple convolutional neural networks. By integrating the Segment Anything Model (SAM) for image content segmentation into specialized scenarios such as sea turtle identification, this includes automating the semantic segmentation of regions of interest in sea turtle images, as well as edge detection and multimodal text comprehension using Contrastive Language-Image Pre-training models. The individual identification is based on the geometric morphology of the facial scales, shell, as well as the fins and neck of sea turtles. This approach allows for precise and detailed identification by leveraging the unique anatomical features of each specimen. The results obtained demonstrate high performance in terms of precision, supported by the evaluation of metrics such as precision and sensitivity. These metrics provide a quantitative measure of the model's ability to accurately identify individuals.

Spatial optical mode sorter is a promising device that can significantly contribute to communication and imaging domains. However, currently available commercial devices fall short in effectively sorting large mode bases with high fidelity, particularly in the visible range. Traditionally, such sorting has relied on multi-plane light conversion (MPLC) setups, wherein the optical field undergoes iterative bouncing between a mirror and a spatial light modulator programmed with different phase profiles. This configuration can be also thought of as a diffractive optical neural network (DONN), where each phase mask acts like a single layer.

In this work we treat the mode sorter as a DONN and instead of employing iterative algorithms to calculate optimal phase masks with conventional MPLC approaches, we propose training it with machine learning algorithms. We investigate various training methodologies, including backpropagation within simulation environments, hybrid approach that incorporate experimental data, and forward-forward training strategies.

Hands play a significant role in human interaction with objects and can enhance our understanding of the activity being performed. They serve as the focal point in egocentric vision. Accurate segmentation of hands from the background and other objects present in a scene enables the understanding of hand movements and gestures. In this paper, we propose an improved hand segmentation algorithm that leverages deep learning techniques to achieve improved segmentation results.

Our approach makes use of optimization techniques that enhance the accuracy of the mask produced using an optimized model (RefineNet-Pix). We employed techniques that help optimize computational resources while improving performance. We then trained the model with the Pascal VOC 2012 dataset and tested the trained model with the Egohands dataset.

The publicly available hands datasets, Egohands, test images was used to evaluate the model performance. Additionally, we compared the result with the baseline model for performance evaluation. The result obtained shows an improved performance of over 5% when compared with the baseline. This enhanced model shows better segmentation accuracy across these metrics (mean intersection over union (mIoU), pixel accuracy (pixelAcc.), and mean accuracy (meanAcc.), and lesser processing time outperforming the baseline model. This improved model can be used in several computer vision applications including object manipulation, activity recognition, rehabilitation therapy, and assistive technologies.

Graph Convolutional Networks (GCNs) offer an innovative method for classifying and identifying segmented images by converting them into superpixels, which serve as nodes in a characteristic graph for each image. This deep learning algorithm generates graph structures with edges representing the distance and color intensity relationship, ensuring accurate representation of image content during training.

GCNs are particularly effective for sea turtle identification due to the unique morphology and patterns of these animals, such as the size, structure, and coloration of their head, carapace, flippers, and characteristic scutes. By leveraging these consistent patterns, GCNs can generate detailed graphs from each turtle, facilitating precise monitoring and research. This includes evaluating geographic distribution, population density, migration, survival, reproduction, and growth of sea turtles.

The proposed GCN algorithm for sea turtle identification is robust and versatile, demonstrating high performance in terms of precision and accuracy. It is resilient to variations in images taken in different environments, angles, and lighting conditions. This marks a significant improvement over other algorithms like Hotspotter by automating the entire process and using segmentation to simplify images and leveraging turtle morphology to generate the graph. Additionally, the use of different color channels allows for the use of images from various environments, reducing the need for specific image characteristics for training. This new algorithm supports the monitoring and protection of sea turtles by providing a powerful tool for their conservation and mapping, contributing significantly to the broader understanding of marine life.

This talk introduces the design and analysis of a passive and fully optical on-chip direction-finding architecture, which can operate effectively in broad or narrow detection scenarios. The proposed system is based on an equally spaced linear array of M grating couplers that steer incoming waves toward a passive and non-unitary photonic processor. This performs a processing operation from which the incident angles of the incoming waves are discretely determined at the N outputs of the photonic processor. The detection functionalities are enhanced by introducing some tracking functions implemented in a post-processing stage at the photonic unit output. The benefits and disadvantages of each tracking function are illustrated. Lastly, an approximation to reduce the photonic unit to a unitary one is discussed, rendering a compact device that operates only on a narrow detection range.

Digital backplane liquid crystal on silicon devices (LCoS) are widely used in wavefront engineering applications due the maturity of the technology, thus the reliability of their performance. A thorough study of the relation between the different control parameters and their effect on the retardance modulation is lacking though. These control parameters, which are the low and high levels of the voltage signal together with the grey level addressed, in relation with the retardance modulation produced are explored across the whole visible spectrum. From a more practical point of view, this enables to decide which voltage parameters across the whole range of possibilities are the most appropriate for a certain wavefront engineering problem without the need to measure them in advance. Additionally, we are interested in the comparison of two different approaches developed in our research group to obtain the retardance modulation to see their possible pros and cons. The results presented are backed by the experimental measurements provided.

We present an encryption system consisting of an electrically tunable metasurface and a matched detector for secure encryption of grayscale images in the 8 – 12 μm wavelength range. In the proposed scheme, the encrypted image corresponds to the spatially varying thermal intensity of the metasurface as captured by its matched detector. Using examples of single and multi-image encryption, we show that the optical properties of either the metasurface or matched detector alone do not reveal any meaningful information about the encrypted image, thereby validating the security of the proposed scheme. The electrical tunability of the metasurface provides an additional layer of security as the image can only be retrieved by operating it at a predefined voltage level. Our results provide intriguing possibilities for the development of compact and secure object tagging and anti-counterfeiting applications in the infrared.

This work presents a methodology for characterizing stellar objects using photometric data obtained from the J-PLUS mission in its third version. The J-PLUS mission does not provide direct astrophysical parameters such as effective temperature and luminosity, which limits our understanding of the stars’ evolutionary state. To overcome this limitation, we perform a data cross-matching with the GAIA DR3 space mission, which offers detailed information on these properties.

By conducting the data cross-matching, we obtain effective temperature and luminosity labels for certain stellar objects in the J-PLUS DR3 catalog. These labels are then utilized as training data for supervised machine learning techniques, specifically the XGBoost algorithm for estimating effective temperature and the Random Forest algorithm for estimating the luminosity of the missing stellar objects in the J-PLUS catalog.

The integration of J-PLUS and GAIA data, combined with the application of supervised learning algorithms, enables accurate and efficient estimation of the missing astrophysical parameters in the J-PLUS DR3 catalog. This methodology significantly contributes to the characterization of the 8,013,455 stars observed by J-PLUS DR3.

Disparity refinement is a post-processing step in stereo vision that retrieves unknown disparity values caused by pixel occlusions or estimation errors. This step is crucial for improving depth estimation accuracy and reducing artifacts. In this work, we propose an iterative method based on genetic optimization to perform disparity refinement for stereo vision. The estimation of unknown disparity values is formulated as an optimization problem, where a fitness function is optimized by minimizing a trade-off between disparity variations and point correspondence errors. The proposed method achieves accurate refined disparity maps for stereo depth estimation. Computer simulation results are presented and discussed in terms of objective performance measures. Additionally, the results are compared with those obtained using a well-known existing method.

Conventional fringe projection profilometers utilize cameras and projectors in the visible spectrum. Nevertheless, some applications require profilometers with a complementary thermal camera for the infrared spectrum. Since the point cloud is computed from pixel correspondences between the visible camera-projector pair, the texture in the visible spectrum is obtained by direct association of color from each image pixel to its corresponding point in the cloud. Unfortunately, the texture from the thermal camera is not straightforward because of the inexistence of pixel-point correspondences. In this paper, a simple interpolation-based method for determining the texture of the reconstructed objects is proposed. The theoretical principles are reviewed, and an experimental verification is conducted using a visible-thermal fringe projection profilometer. This work provides a helpful framework for three-dimensional data fusion for advanced multi-modal profilometers.

Binocular vision is an effective technique for depth estimation in outdoor applications. However, the performance of this technique can be affected by low-contrasted images captured in the presence of scattering particles. This paper presents a binocular vision-based method for image dehazing and depth estimation in a scattering medium. First, the disparity map of the scene is computed from captured binocular images. Next, the atmospheric parameters of the medium and scene depth are determined with a proposed estimation method. Finally, undegraded images of the scene are obtained using an atmospheric optics restoration method. The theoretical foundations of the proposed approach are reviewed, and an experimental validation is presented utilizing a laboratory platform.

This paper presents the development of an autonomous navigation system for Unmanned Aerial Vehicles (UAVs) using visual reference. The proposal employs a Convolutional Neural Network (CNN) to classify traffic signal images, enabling UAVs to navigate evolving dynamic environments. This research involves the configuration of the Robot Operating System (ROS) for UAV communication, the implementing of a specialized CNN for image classification, and the integration of this network into the navigation system. Therefore, a system will be presented for image acquisition and UAV manipulation based on CNN outputs. We present experimental results specially designed to demonstrate the efficiency of the proposal, to validate the analysis and implementation.

This work presents the development of an autonomous mapping and navigation system tailored for a holonomic robot operating in unknown environments, leveraging Light Detection and Ranging (LiDAR) technology. The core of this research lies in integrating a sophisticated Simultaneous Localization and Mapping (SLAM) algorithm with real-time LiDAR data processing and the GPU’s parallel capabilities to process several subspaces concurrently. This work introduces a novel method to accelerate route planning in an omnidirectional mobile robot fusing advances in hardware with sophisticated algorithmic techniques, a new paradigm is established in path planning for omnidirectional mobile robots, marking an important milestone in the search for more agile and capable robotic systems.

Integrating quantum algorithms with machine and deep learning models has emerged as a promising method for addressing medical image classification challenges. This integration can enhance speed and efficiency when performing complex computations. However, hybrid quantum models, particularly on Quantum Convolutional Neural Networks (QCNNs) face two significant drawbacks: the placement of the quantum convolutional layer before the model architecture and the lack of integration of the quantum layer within the training process. These disadvantages reduce the robustness and reproducibility of the models. This study proposes that integrates the quantum layer into the quantum layer to address these shortcomings. We present a comparative analysis between a hybrid quantum deep learning model, which includes a trainable quantum layer, and its classical counterpart for the classification of skin cancer dermatoscopic images. The hybrid model attains 0.7865 of accuracy, a recall of 0.7321, a precision of 0.7268, and an F1 Score of 0.7288, while the classical model reaches an accuracy, recall, precision, and F1 Score of 0.8510, 0.8472, 0.8495, and 0.8447. The hybrid model achieves comparable results to its classical counterpart and demonstrates the advantages of weight adjustment in quantum layers and their potential in improving medical imaging analysis.

The colorization of monochromatic images has demonstrated utility in enhancing human comprehension of images and boosting the accuracy of succeeding image-processing tasks. Nonetheless, current fully automated colorization methodologies often exhibit optimal performance based on the input image’s nature and the employed algorithms’ architectural specifics. In response to this challenge, this paper introduces a novel methodology aimed at effectively predicting the most suitable colorization model for a given input image. This comprehensive approach is characterized by exceptional accuracy across diverse datasets.

A crucial task in facial expression recognition is the classification of facial features in captured images. This classification task is challenging because facial features change dynamically due to several facial expressions. Additionally, the captured face images are often degraded by additive noise, nonuniform illumination, geometrical modifications, and partial occlusions, increasing uncertainty in classification . Several successful methods for facial landmark classification based on machine learning have been proposed. This work presents a comparative study of existing classification methods for facial landmarks in image sequences degraded by noise, nonuniform illumination, and partial occlusions. The performance of the classification methods considered in the study is quantified in terms of accuracy using face images from well-known datasets. The study aims to provide useful insights into the efficacy of existing facial landmark classification methods under challenging conditions.

Accurate individual identification of animals plays a crucial role in population studies, providing invaluable insights for species research and conservation researchers. Such identification enables the tracking of migratory patterns, growth rates, and survival rates, which are essential for understanding species dynamics and implementing effective conservation strategies.

In this work presents an AI-based approach for individual sea turtle identification using Photo-ID. Comparing an own-developed algorithm with the HotSpotter ID, a well-known method for semi-automated identification of individual sea turtles within populations, highlighting differences in image preprocessing and One-vs-Many matching evaluation. Tested on a diverse database of sea turtle images. The database images vary in environments, light conditions, qualities and resolutions, at different distances, and angles. The proposed Photo-ID method involves automatic detection of the turtle’s head using a modified YOLOv5 network, extraction of pixels to emphasize head scale patterns, and the ORB algorithm for identification. With standard quality images, the proposed algorithm achieves 98% accuracy and add an end-to-end pipeline to the photoidentification process, outperforming the HotSpotter’s 96% and required for human intervention. This approach enhances the identification process, contributing to endangered species protection and addressing the challenges associated with the time-consuming nature of traditional Photo-ID methods.

Hyperthermia is a cancer treatment that utilizes magnetic nanofluidic and magnetic induction. The experimental evaluation techniques used to validate heat transfer and changes in thermal diffusivity include the use of phantom models to simulate human tissue and infrared thermography. The present study applies phantom models with magnetic inserts, which pose challenges in determining temperature distribution due to their location and depth, affect thermal contrast. Processing thermograms is crucial for identifying areas of interest, which enables the development of new hyperthermia evaluation techniques.

Spatial Light Modulators (SLMs) typically use PA-LCoS microdisplays, known for high resolution and small pixel size, for phase-only modulation. However, different phenomena related to the cross-talk, fringing fields between pixels, and out-of-plane reorientation of the liquid crystal director degrade the performance of pixelated SLM devices. The technology faces the challenge of overcoming these phenomena as novel applications require higher pixel size resolutions and smaller pixel sizes. This work applies a numerical workflow that permits the analysis of the different parameters that set up a PA-LCoS microdisplay, e.g., fill factor, pixel size, external voltage and the influence of the LC director and the interaction of light through the structure. Specifically, we focus our analysis on the impact of high-frequency binary phase gratings on the radiometric and polarimetric response with orthogonal grating vectors. The heterogenous LC distribution as a function of the grey level along the 3D structure and the inhomogeneous distribution of the Stokes parameters are analysed, setting up a framework to analyse the nonlinear behaviour of the system and its impact on diffraction and polarimetric performance in future microdisplays with higher resolutions and smaller pixel sizes.

Traditional fluorescence microscopes, limited in resolution, impede the precise identification of subcellular compartments. While super-resolution microscopes are frequently employed to compensate for this deficiency, their application in multicell or live tissue studies poses challenges due to inherent tradeoffs and high expenses. On the other side, understanding the functionality of living cells and exploring intracellular dynamics within their natural organismic environment demands sophisticated and costly equipment, which may not be affordable for all laboratories. In this investigation, we applied computational methods, specifically employed super-resolution radial fluctuations (SRRF), on the mouse thigh to capture sequences of tissue micrographs. By systematically exploring various numerical modifications and parameters, we identified specific factors that significantly enhanced the resolution of subcellular structures.

The subcellular imaging in the thigh muscle was conducted for this purpose. Standard imaging protocols encompassed capturing sequences of image series with varying frame counts at different rates, utilizing high numerical aperture (NA) objective lenses to explore multiple parameters for optimal results. Subsequently, following sequential numerical adjustments to minimize background noise and enhance signal intensity, the SRRF algorithm was employed on the image stacks. The result was frames of muscle fibers with significantly improved resolution. In the final images, discrete organelle structures and dynamics were discernible, overcoming the poor lateral resolution of the original microscope images that depicted indistinct schematics of organelles. Importantly, this technique is versatile, requiring neither specific systems nor components, and it does not entail additional costs.

A microscope design was developed using 3D printing and liquid crystals to achieve variable polarization configurations and semi-automatic processes without the need for mechanical movement of polarizers and retarders. The device includes a polariscope composed of electronically controlled liquid crystals and a compliant mechanism for precise micrometric movements. Stepper motors and software control these movements, and the process of acquiring images requires connecting a sensor to a Raspberry Pi. The microscope's lens and extension tube assembly magnify the image 600 times. The study compares two polarization techniques: Stokes and ellipsometric polarization. Additionally, it analyzes crystals and meiotic spindles of maturing porcine oocytes. The ellipsometric technique is more effective in detecting low retards, as indicated by the results. This prototype has the potential to reduce the cost of in vitro fertilization in assisted reproduction laboratories, with a focus on animals of high genetic value.

This paper presents the development of a LiDAR-based object classification system using machine learning and signal processing. The proposal explores Support Vector Machines (SVM) and neural networks to classify terrain with the help of a LiDAR that scans an area similarly to how a picture is taken. This project involves the processing of data to generate a point cloud that lets us visualize the scans taken by the Light Detection and Ranging (LiDAR). The dataset was built by taking multiple scans of three types of terrain, flat, grassy, and rocky. This paper shows experimental results of machine learning models built around LiDAR-acquired data and small datasets, it also shows point cloud visualizations and a simple signal processing technique.

This work presents a novel fiber optic Mach-Zehnder interferometer (MZI) sensor that can measure curvature due to its structure. This structure employs a 2-section MZI filter constructed from SMF-28 fiber using the core offset technique. The source in the 1480-1600 nm range shows intermodal energy interference between the core and the cladding and exhibits six spacing notches for sensing applications. The sensibility obtained was of 0.00018 nm/uW when curvature is applied. However, by including a thin film based on Zinc Oxide (ZnO) and another with Aluminum-doped zinc oxide (AZO) in the sensing arrangement, an increase in the sensitivity was detected with a value of 0.00015 nm/uW and 0.00016 nm/uW, respectively. This sensor can be used for applications in various fields, such as environmental monitoring, engineering, and process control.

In the last years many efforts have been invested in the development of augmented reality devices. Depending on the application different constraints need to be faced. Providing full color see-through augmented reality on eyeglasses with ophthalmic correction, and valid for a wide range of use cases, is one of the most challenging and ambitious applications. We are within a European Project aiming to this goal. One of the key components in this eyewear is the holographic lens mirror (HLM), acting as the beam combiner responsible for the see-through capability. In this work, we present the main goals of this European Project and, more specifically, the holographic approach being developed for the HLM within our research group.

Aperture masking observations for binary stars have been done with a 1.56-m telescope at Shanghai Astronomical Observatory(SHAO) during 2019-2024. In order to ensure ample light available to restore the high resolution images even at short exposure times and reach the diffraction limit of the 1.56-m telescope, we select some binary stars and reference stars nearby with magnitude 5-8 and angular distance 0.2-4 arcsec in the WDS catalog as observation targets. This article achieved high-resolution restoration of binary stars using a hybrid data processing method, including data reduction, a spatial domain method named ISA or a frequency domain method named SI to suppress atmospheric turbulence, and OS-EM to reduce image degradation caused by multi-aperture interference. The results show that this method can effectively obtain high-resolution images of binary stars, and the measured angular distance is basically consistent with the given value in WDS catalog.

The first demonstration of indirect drive fusion in the National Ignition Facility (NIF) has ushered in a new era of laser-based fusion work. Before every experiment in NIF all 192 laser beams passing through various laser sub-sections must be aligned to the target chamber center within 30 minutes with a 50-micron accuracy at the target chamber center. NIF uses a CCD camera-based imaging system to identify the beam location; the laser beams are then aligned using motorized mirrors within a feedback control system. The success of the automatic alignment (AA) is contingent upon calculating the exact location of the laser beam in the CCD image of the beam. When an algorithm fails due to poor imaging, gradient illumination, or high noise content, the automated operation for the affected alignment loop stops, and the alignment enters a manual mode with operator intervention. The operator/shot director must decide if the imaging condition needs to be improved to successfully pass alignment or if the beam has to be dropped from participating in the shot thus reducing the total energy delivered to the target. To minimize such failures where a legitimate but challenging beam image is present, additional alternate algorithms are added to the original algorithm. When the first algorithm fails, this alternate algorithm is executed to process the image even though it may have a slightly higher uncertainty. Although the second algorithm may report a higher uncertainty, this system design allows the loop to continue to align the laser without operator intervention. The implication of this approach may be far-reaching, for example, in the application of driver-less cars, where the original algorithm developed by training on a large set of data, may sometimes fail to account for the presence of a pedestrian and lead to a fatal accident. Here we hypothesize that a back-up algorithm based on a second approach will minimize such risks in the real-world driver-less driving scenarios.