Neural Network-based Algorithm Research Articles

Bubble Detection in Multiphase Flows Through Computer Vision and Deep Learning for Applied Modeling

An innovative method for bubble detection and characterization in multiphase flows using advanced computer vision and neural network algorithms is introduced. Building on the research group’s previous findings, this study combines high-speed video capture with advanced deep learning techniques to enhance bubble detection accuracy and dynamic analysis. In order to further develop a robust framework for detecting and analyzing bubble properties in multiphase flows, enabling accurate estimation of essential mass transfer parameters, a YOLOv9-based neural network was implemented for bubble segmentation and trajectory analysis, achieving high accuracy. Key contributions include the development of an averaged mass transfer model integrating experimental data, neural network outputs, and scaling algorithms, as well as validation of the proposed methodology through experimental studies, including high-speed video imaging and comparisons with mass transfer coefficients obtained via the sulfite method. By precisely characterizing critical parameters, the algorithm enables accurate gas transfer rate calculations, ensuring optimal conditions in various industrial applications. The neural network-based algorithm serves as a non-invasive platform for detailed characterization of bubble media, demonstrating high accuracy in experimental validation and significantly outperforming traditional techniques. This approach provides a robust tool for real-time monitoring and modeling of bubble flows, laying the foundation for novel, non-invasive methods to measure multiphase media properties.

Advanced Hybrid Neural Networks for Accurate Recognition of the Extended Alphabet and Dynamic Signs in Mexican Sign Language (MSL)

The Mexican deaf community primarily uses Mexican Sign Language (MSL) for communication, but significant barriers arise when interacting with hearing individuals unfamiliar with the language. Learning MSL requires a substantial commitment of at least 18 months, which is often impractical for many hearing people. To address this gap, we present an MSL-to-Spanish translation system that facilitates communication through a spelling-based approach, enabling deaf individuals to convey any idea while simplifying the AI’s task by limiting the number of signs to be recognized. Unlike previous systems that focus exclusively on static signs for individual letters, our solution incorporates dynamic signs, such as “k”, “rr”, and “ll”, to better capture the nuances of MSL and enhance expressiveness. The proposed Hybrid Neural Network-based algorithm integrates these dynamic elements effectively, achieving an F1 score of 90.91%, precision of 91.25%, recall of 91.05%, and accuracy of 91.09% in the extended alphabet classification. These results demonstrate the system’s potential to improve accessibility and inclusivity for the Mexican deaf community.

Implementation of Smart Farm Systems Based on Fog Computing in Artificial Intelligence of Things Environments.

Cloud computing has recently gained widespread attention owing to its use in applications involving the Internet of Things (IoT). However, the transmission of massive volumes of data to a cloud server often results in overhead. Fog computing has emerged as a viable solution to address this issue. This study implements an Artificial Intelligence of Things (AIoT) system based on fog computing on a smart farm. Three experiments are conducted to evaluate the performance of the AIoT system. First, network traffic volumes between systems employing and not employing fog computing are compared. Second, the performance of the communication protocols-hypertext transport protocol (HTTP), message queuing telemetry transport protocol (MQTT), and constrained application protocol (CoAP)-commonly used in IoT applications is assessed. Finally, a convolutional neural network-based algorithm is introduced to determine the maturity level of coffee tree images. Experimental data are collected over ten days from a coffee tree farm in the Republic of Korea. Notably, the fog computing system demonstrates a 26% reduction in the cumulative data volume compared with a non-fog system. MQTT exhibits stable results in terms of the data volume and loss rate. Additionally, the maturity level determination algorithm performed on coffee fruits provides reliable results.

Development and validation of a deep learning-based survival prediction model for pediatric glioma patients: A retrospective study using the SEER database and Chinese data

ObjectiveDevelop a time-dependent deep learning model to accurately predict the prognosis of pediatric glioma patients, which can assist clinicians in making precise treatment decisions and reducing patient risk. Study designThe study involved pediatric glioma patients from the Surveillance, Epidemiology, and End Results (SEER) Registry (2000–2018) and Tangdu Hospital in China (2010–2018) within specific time frames. For training, we selected two neural network-based algorithms (DeepSurv, neural multi-task logistic regression [N-MTLR]) and one ensemble learning-based algorithm (random survival forest [RSF]). Additionally, a multivariable Cox proportional hazard (CoxPH) model was developed for comparison purposes. The SEER dataset was randomly divided into 80 % for training and 20 % for testing, while the Tangdu Hospital dataset served as an external validation cohort. Super-parameters were fine-tuned through 1000 repeated random searches and 5-fold cross-validation on the training cohort. Model performance was assessed using the concordance index (C-index), Brier score, and Integrated Brier Score (IBS). Furthermore, the accuracy of predicting survival at 1, 3, and 5 years was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and the area under the ROC curves (AUC). The generalization ability of the model was assessed using the C-index of the Tangdu Hospital data, ROC curves for 1, 3, and 5 years, and AUC values. Lastly, decision curve analysis (DCA) curves for 1, 3, and 5-year time frames are provided to assess the net benefits across different models. ResultsA total of 9532 patients with pediatric glioma were included in this study, comprising 9274 patients from the SEER database and 258 patients from Tangdu Hospital in China. The average age at diagnosis was 9.4 ± 6.2 years, and the average survival time was 96 ± 66 months. Through comprehensive performance comparison, the DeepSurv model demonstrated the highest effectiveness, with a C-index of 0.881 on the training cohort. Furthermore, it exhibited excellent accuracy in predicting the 1-year, 3-year, and 5-year survival rates (AUC: 0.903–0.939). Notably, the DeepSurv model also achieved remarkable performance and accuracy on the Chinese dataset (C-index: 0.782, AUC: 0.761–0.852). Comprehensive analysis of DeepSurv, N-MTLR, and RSF revealed that tumor stage, radiotherapy, histological type, tumor size, chemotherapy, age, and surgical method are all significant factors influencing the prognosis of pediatric glioma. Finally, an online version of the pediatric glioma survival predictor based on the DeepSurv model has been established and can be accessed through https://pediatricglioma-tangdu.streamlit.app. ConclusionsThe DeepSurv model exhibits exceptional efficacy in predicting the survival of pediatric glioma patients, demonstrating strong performance in discrimination, calibration, stability, and generalization. By utilizing the online version of the pediatric glioma survival predictor, which is based on the DeepSurv model, clinicians can accurately predict patient survival and offer personalized treatment options.

Deep learning for solving initial path optimization of mean-field systems with memory

We consider the problem of finding the optimal initial investment strategy for a system modelled by a linear McKean–Vlasov (mean-field) stochastic differential equation with delay, driven by Brownian motion and a pure jump Poisson random measure. The goal is to determine the optimal initial values for the system in the period [ − δ , 0 ] , where δ > 0 is a delay constant, before the system starts at t = 0. Due to the delay in the dynamics, the system will, after startup, be influenced by these initial investment values. It is known that linear stochastic delay differential equations are equivalent to stochastic Volterra integral equations. By utilizing this equivalence, we can find implicit expressions for the optimal investment. Moreover, we propose a deep neural network-based algorithm to solve the stochastic control problem with delay. Specifically, we employ a multi-layer feed-forward neural network for control modelling in the interval [ − δ , 0 ] , and use back-propagation to train the feed-forward neural network. The gradient of the loss function is computed using stochastic gradient descent (SGD) with respect to the weights of the network.

ECAD-Net: Editable Parametric CAD Models Reconstruction from Dumb B-Rep Models Using Deep Neural Networks

This paper introduces a novel framework capable of reconstructing editable parametric CAD models from dumb B-Rep models. First, each B-Rep model is represented with a network-friendly formalism based on UV-graph, which is then used as input of eCAD-Net, the new deep neural network-based algorithm that predicts feature-based CAD modeling sequences from the graph. Then, the sequences are scaled and fine-tuned using a feature matching algorithm that retrieves the exact parameter values from the input dumb CAD model. The output sequences are then converted in a series of CAD modeling operations to create an editable parametric CAD model in any CAD modeler. A cleaned dataset is used to learn and validate the proposed approach, and is provided with the article. The experimental results show that our approach outperforms existing methods on such reconstruction tasks, and it outputs editable parametric CAD models compatible with existing CAD modelers and ready for use in downstream engineering applications.

Lightweight deep neural network for radio frequency interference detection and segmentation in synthetic aperture radar

Radio frequency interference (RFI) poses challenges in the analysis of synthetic aperture radar (SAR) images. Existing RFI suppression systems rely on prior knowledge of the presence of RFI. This paper proposes a lightweight neural network-based algorithm for detecting and segmenting RFI (LDNet) in the time-frequency domain. The network accurately delineates RFI pixel regions in time-frequency spectrograms. To mitigate the impact on the operational speed of the entire RFI suppression system, lightweight modules and pruning operations are introduced. Compared to threshold-based RFI detection algorithms, deep learning-based segmentation networks, and AC-UNet specifically designed for RFI detection, LDNet achieves improvements in mean intersection over union (MIoU) by 24.56%, 13.29%, and 7.54%, respectively.Furthermore, LDNet reduces model size by 99.03% and inference latency by 24.53% compared to AC-UNet.

Multimodal Convolutional Neural Network-Based Algorithm for Real-Time Detection and Differentiation of Malignant and Inflammatory Biliary Strictures in Cholangioscopy: a Proof of Concept Study with Video

Deep learning algorithms gained attention for detection (computer-aided detection [CADe]) of biliary tract cancer in digital single-operator cholangioscopy (dSOC). We developed a multimodal convolutional neural network (CNN) for detection (CADe), characterization and discriminating (computer-aided diagnosis [CADx]) between malignant, inflammatory, and normal biliary tissue in raw dSOC videos. In addition, clinical metadata were included in the CNN algorithm to overcome limitations of image-only models. Based on dSOC videos and images of 111 patients (total of 15,158 still frames), a real-time CNN-based algorithm for CADe and CADx was developed and validated. We established an image-only model and metadata injection approach. In addition, frame-wise and case-based predictions on complete dSOC video sequences were validated. Model embeddings were visualized, and class activation maps highlighted relevant image regions. The concatenation-based CADx approach achieved a per-frame area under the receiver-operating characteristic curve of .871, sensitivity of .809 (95% CI, .784-.832), specificity of .773 (95% CI, .761-.785), positive predictive value of .450 (95% CI, .423-.467), and negative predictive value of .946 (95% CI, .940-.954) with respect to malignancy on 5715 test frames from complete videos of 20 patients. For case-based diagnosis using average prediction scores, 6 of 8 malignant cases and all 12 benign cases were identified correctly. Our algorithm distinguishes malignant and inflammatory bile duct lesions in dSOC videos, indicating the potential of CNN-based diagnostic support systems for both CADe and CADx. The integration of non-image data can improve CNN-based support systems, targeting current challenges in the assessment of biliary strictures.

Hybrid Long Short-Term Memory Fused Convolution Neural Network for Weather Forecasting

Weather forecasting is an effort by meteorologists to predict whether states at a few prospective times and the conditions of weather that might be estimated. With new modern technology, present weather forecasting methods are highly precise. To achieve high accuracy, the methods developed for weather forecasting were much more complicated owing to many factors. Here, the usage of time series data for weather forecasting is done by the devised Long-Short Term Memory fused Convolutional neural network (LSTMFCNN). At first, the acquisition of input time series data from the specific dataset is done. In The feature extraction technical features are done by considering the input time series data. Then, the feature extraction is done utilizing the Rider Optimization Algorithm-Based Neural Network (RideNN) with the Soergel metric. RideNN is the integration of the Rider Optimization Algorithm (ROA) with the Neural Network (NN) classifier. Thus, the feature fusion step reduces the complexity and improves the accuracy. Thereafter, the oversampling technique is utilized for the data augmentation (DA) process. Finally, weather forecasting is done utilizing the newly designed LSTMFCNN and is obtained by the integration of Convolutional Neural Network (CNN) and Deep Long-Short Term Memory (DLSTM).

Achieving Image Encryption Quantum Dot-Functionalized Encryption Camera with Designed Films.

The risk of information leaks increases as images become a crucial medium for information sharing. There is a great need to further develop the versatility of image encryption technology to protect confidential and sensitive information. Herein, using high spatial redundancy (strong correlation of neighboring pixels) of the image and the in situ encryption function of a quantum dot functionalized encryption camera, in situ image encryption is achieved by designing quantum dot films (size, color, and full width at half maximum) to modify the correlation and reduce spatial redundancy of the captured image during encryption processing. The correlation coefficients of simulated encrypted image closely apporach to 0. High-quality decrypted images are achieved with a PSNR of more than 35dB by a convolutional neural network-based algorithm that meets the resolution requirements of human visual perception. Compared with the traditional image encryption algorithms, chaotic image encryption algorithms and neural network-based encryption algorithms described previously, it provides a universal, efficient and effective in situ image encryption method.

An overview of hand gesture recognition based on computer vision

Hand gesture recognition emerges as one of the foremost sectors which has gone through several developments within pattern recognition. Numerous studies and research endeavors have explored methodologies grounded in computer vision within this domain. Despite extensive research endeavors, there is still a need for a more thorough evaluation of the efficiency of various methods in different environments along with the challenges encountered during the application of these methods. The focal point of this paper is the comparison of different research in the domain of vision-based hand gesture recognition. The objective is to find out the most prominent methods by reviewing efficiency. Concurrently, the paper delves into presenting potential solutions for challenges faced in different research. A comparative analysis particularly centered around traditional methods and convolutional neural networks like random forest, long short-term memory (LSTM), heatmap, and you only look once (YOLO). considering their efficacy. Where convolutional neural network-based algorithms performed best for recognizing the gestures and gave effective solutions for the challenges faced by the researchers. In essence, the findings of this review paper aim to contribute to future implementations and the discovery of more efficient approaches in the gesture recognition sector.

Practical Application and Case Analysis of Computer Image Vision Technology in Music Education and Teaching

Exploring how to utilize images to enrich music teaching content and provide a more visually impactful learning experience is an important topic. Therefore, this paper introduces a convolutional neural network-based algorithm for extracting audio features to construct a music visualization model. By identifying features such as note pitches, it enhances pitch recognition and integrates with CNN algorithms for audio information visualization. Experimental results demonstrate an accuracy rate exceeding 97%, showcasing the significant advantage of this method in visualizing audio information in music multimedia classrooms. It provides technical support for bringing a new visual experience to music education.

Neural network-based algorithm for door handle recognition using RGBD cameras

The ability to recognize and interact with a variety of doorknob designs is an important component on the path to true robot adaptability, allowing robotic systems to effectively interact with a variety of environments and objects The problem addressed in this paper is to develop and implement a method for recognizing the position of a door handle by a robot using data from an RGBD camera. To achieve this goal, we propose a revolutionary approach designed for autonomous robots that allows them to identify and manipulate door handles in different environments using data obtained from RGBD cameras. This was achieved by creating and annotating a complete dataset consisting of 5000 images of door handles from different angles, with the coordinates of the vertices of the bounding rectangles labeled. The architectural basis of the proposed approach is based on MobileNetV2, combined with a special decoder that optimally increases the resolution to 448 pixels. A new activation function specially designed for this neural network is implemented to ensure increased accuracy and efficiency of raw data processing. The most important achievement of this study is the model's ability to work in real-time, processing up to 16 images per second. This research paves the way for new advancements in the fields of robotics and computer vision, making a substantial contribution to the practical deployment of autonomous robots in a myriad of life's spheres.

MOTOR IMAGERY SIGNAL CLASSIFICATION FOR BRAIN–COMPUTER INTERFACE USING RideNN WITH HOLO-ENTROPY FEATURES

The brain–computer interface (BCI) database’s motor assessment depends heavily on the motor imagery (MI) signal classification. By examining the multiple patterns of different creative tasks in the electroencephalogram (EEG) signals, the intention of humans is translated into computer-based commands in the MI-based signals of BCI data. Nevertheless, low accuracy and efficiency are issues with MI–EEG signals’ classification because of the low signal-to-noise ratio, huge individual differences, overall volatility, and complexity in the signal. To overcome these problems, this research proposes a rider optimization algorithm-based neural network (ROA-based NN) to classify the MI signals effectively. Pre-processing is done after collecting the dataset of raw EEG signals. The suitable electrodes, such as C3, C4, and Cz, are subsequently chosen from the MI signals. Using the holo-entropy-based WPD feature extractor, tunable Q-factor wavelet transform (T-QWT), and common spatial patterns in the model, the pertinent features are extracted from the chosen electrodes. The developed holo-entropy-based WPD feature examines the electrode structure’s association. As a result, the most diverse signals are removed from the chosen electrodes before being input into the proposed RideNN classifier, where the proposed ride optimization algorithm optimizes classification performance and correctly predicts and classes the output from the MI signals that have been analyzed. The developed RideNN classifier recognizes the patterns more accurately processes more data and tackles the noise and incomplete data effectively. Utilizing the parameters of accuracy, sensitivity, and specificity, the results are evaluated. The PROA-based RideNN and the combined features obtain the maximum accuracy of 92.24%, sensitivity of 92.26%, and specificity of 92.14% for the BCI competition-IV 2a database. The qPROA-based RideNN and the combined features obtain the maximum accuracy of 92.11%, sensitivity of 91.98%, and specificity of 92.35% for the BCI competition-IV 2b database.

Improving Earth-like planet detection in radial velocity using deep learning

Context. Many novel methods have been proposed to mitigate stellar activity for exoplanet detection as the presence of stellar activity in radial velocity (RV) measurements is the current major limitation. Unlike traditional methods that model stellar activity in the RV domain, more methods are moving in the direction of disentangling stellar activity at the spectral level. As deep neural networks have already been proven to be one of the most effective tools in data mining, in this work, we explore their potential in the context of Earth-like planet detection in RV measurements. Aims. The goal of this paper is to present a novel convolutional neural network-based algorithm that efficiently models stellar activity signals at the spectral level, enhancing the detection of Earth-like planets. Methods. Based on the idea that the presence of planets can only produce a Doppler shift at the spectral level while the presence of stellar activity can introduce a variation in the profile of spectral lines (asymmetry and depth change), we trained a convolutional neural network to build the correlation between the change in the spectral line profile and the corresponding RV, full width at half maximum (FWHM) and bisector span (BIS) values derived from the classical cross-correlation function. Results. This algorithm has been tested on three intensively observed stars: Alpha Centauri B (HD 128621), Tau ceti (HD 10700), and the Sun. By injecting simulated planetary signals at the spectral level, we demonstrate that our machine learning algorithm can achieve, for HD 128621 and HD 10700, a detection threshold of 0.5 m s−1 in semi-amplitude for planets with periods ranging from 10 to 300 days. This threshold would correspond to the detection of a ~4 M⊕ in the habitable zone of those stars. On the HARPS-N solar dataset, thanks to significantly more data, our algorithm is even more efficient at mitigating stellar activity signals and can reach a threshold of 0.2 m s−1, which would correspond to a 2.2 M⊕ planet on the orbit of the Earth. Conclusions. To the best of our knowledge, it is the first time that such low detection thresholds are reported for the Sun, but also for other stars, and therefore this highlights the efficiency of our convolutional neural network-based algorithm at mitigating stellar activity in RV measurements.

On-Device Neural Network for Object Train and Recognition using Mobile

A neural network is a machine learning (ML) program or model that processes information and recognizes patterns, similar to the human brain. The neural network algorithm operates by training on data to learn and enhance its accuracy. If there is any mistake in the learning process, the machine will react unexpectedly and produce incorrect information. So, whenever a neural network model is developed, it is mandatory to evaluate its performance and ensure its output is accurate. Cameras, touchscreens, internet connectivity, and powerful CPUs have contributed to the popularity of smartphones. Mobile apps are software applications that run on smartphones. ML enhances mobile app functionality by offering features such as voice recognition, image analysis, natural language processing, personalization, and recommendations. Training ML models using mobile apps is challenging due to limited resources, data, and privacy concerns. Object recognition is a neural network-based technique that enables the identification and localization of objects in an image or video. This technique is used in driverless cars, disease identification, industrial inspection, robotics, and more. In this paper, the author introduces a new neural network-based algorithm that utilizes mobile devices for training and recognizing images. Although mobile devices have some technological limitations for training, well-established guidelines for systematically mapping verification and validation techniques have been used in the proposed neural network to ensure performance and correctness in on-device object training and recognition.

Implementation of a Long Short-Term Memory Neural Network-Based Algorithm for Dynamic Obstacle Avoidance.

Autonomous mobile robots are essential to the industry, and human-robot interactions are becoming more common nowadays. These interactions require that the robots navigate scenarios with static and dynamic obstacles in a safely manner, avoiding collisions. This paper presents a physical implementation of a method for dynamic obstacle avoidance using a long short-term memory (LSTM) neural network that obtains information from the mobile robot's LiDAR for it to be capable of navigating through scenarios with static and dynamic obstacles while avoiding collisions and reaching its goal. The model is implemented using a TurtleBot3 mobile robot within an OptiTrack motion capture (MoCap) system for obtaining its position at any given time. The user operates the robot through these scenarios, recording its LiDAR readings, target point, position inside the MoCap system, and its linear and angular velocities, all of which serve as the input for the LSTM network. The model is trained on data from multiple user-operated trajectories across five different scenarios, outputting the linear and angular velocities for the mobile robot. Physical experiments prove that the model is successful in allowing the mobile robot to reach the target point in each scenario while avoiding the dynamic obstacle, with a validation accuracy of 98.02%.

Development and validation of machine learning models to predict unplanned hospitalizations of patients with diabetes within the next 12 months

BACKGROUND: The incidence of diabetes mellitus (DM) both in the Russian Federation and in the world has been steadily increasing for several decades. Stable population growth and current epidemiological characteristics of DM lead to enormous economic costs and significant social losses throughout the world. The disease often progresses with the development of specific complications, while significantly increasing the likelihood of hospitalization. The creation and inference of a machine learning model for predicting hospitalizations of patients with DM to an inpatient medical facility will make it possible to personalize the provision of medical care and optimize the load on the entire healthcare system.AIM: Development and validation of models for predicting unplanned hospitalizations of patients with diabetes due to the disease itself and its complications using machine learning algorithms and data from real clinical practice.MATERIALS AND METHODS: 170,141 depersonalized electronic health records of 23,742 diabetic patients were included in the study. Anamnestic, constitutional, clinical, instrumental and laboratory data, widely used in routine medical practice, were considered as potential predictors, a total of 33 signs. Logistic regression (LR), gradient boosting methods (LightGBM, XGBoost, CatBoost), decision tree-based methods (RandomForest and ExtraTrees), and a neural network-based algorithm (Multi-layer Perceptron) were compared. External validation was performed on the data of the separate region of Russian Federation.RESULTS: The best results and stability to external validation data were shown by the LightGBM model with an AUC of 0.818 (95% CI 0.802–0.834) in internal testing and 0.802 (95% CI 0.773–0.832) in external validation.CONCLUSION: The metrics of the best model were superior to previously published studies. The results of external validation showed the relative stability of the model to new data from another region, that reflects the possibility of the model’s application in real clinical practice.

A novel hybrid optimized incremental relevance vector machine and filtering technique for state of charge estimation of lithium-ion batteries

Despite not requiring a comprehensive understanding of the intricate electrochemical reactions within the battery, the data-driven approach faces challenges regarding state of charge (SOC) fluctuations during abrupt changes in battery current. Currently, research on data-driven filtering methods predominantly focuses on neural network-based algorithms, with limited exploration of the relevance vector machine (RVM) algorithm that possesses advantages in handling small sample sizes, high sparsity, and probabilistic distributions. Drawing inspiration from this premise, the present manuscript introduces an innovative fusion algorithm, founded upon the framework of the adaptive Kalman filtering (AKF) algorithm, which intricately melds the optimized incremental RVM (OIRVM) with the Coulomb counting methodology, called Coulomb_AKF_OIRVM (CFR) algorithm. Experimental findings suggest that the CFR algorithm displays exceptional generalization capacity across different driving cycles and temperatures, exhibiting high estimation accuracy and robustness. The average root mean square error, determination coefficient, mean absolute error and maximum absolute error of the public datasets and experimental datasets are 0.25 %, 0.02 %, 0.20 %, and 0.85 %, respectively.

Computational 3D topographic microscopy from terabytes of data per sample

We present a large-scale computational 3D topographic microscope that enables 6-gigapixel profilometric 3D imaging at micron-scale resolution across >110 cm2 areas over multi-millimeter axial ranges. Our computational microscope, termed STARCAM (Scanning Topographic All-in-focus Reconstruction with a Computational Array Microscope), features a parallelized, 54-camera architecture with 3-axis translation to capture, for each sample of interest, a multi-dimensional, 2.1-terabyte (TB) dataset, consisting of a total of 224,640 9.4-megapixel images. We developed a self-supervised neural network-based algorithm for 3D reconstruction and stitching that jointly estimates an all-in-focus photometric composite and 3D height map across the entire field of view, using multi-view stereo information and image sharpness as a focal metric. The memory-efficient, compressed differentiable representation offered by the neural network effectively enables joint participation of the entire multi-TB dataset during the reconstruction process. Validation experiments on gauge blocks demonstrate a profilometric precision and accuracy of 10 µm or better. To demonstrate the broad utility of our new computational microscope, we applied STARCAM to a variety of decimeter-scale objects, with applications ranging from cultural heritage to industrial inspection.