Search Results (8,277)

Fatigue during table tennis training can impair performance and increase the risk of injury. This study proposes a real-time fatigue recognition system that integrates an embedded edge-AI racket with a camera for synchronized multimodal data acquisition. On the racket side, a neural network was deployed on an edge-AI microcontroller unit (MCU) to enable on-device inference, while the test data were transmitted to a computer for multimodal feature fusion. Given the continuous nature of fatigue development, the data were divided into three stages: non-fatigue, transition, and fatigue to facilitate progressive evaluation. In the binary classification task, which distinguished between the two extreme states, the multimodal hybrid model achieved an accuracy of 99%. When the transition stage was incorporated into a three-class classification task, performance declined because of the ambiguous boundaries between fatigue states. However, the use of multimodal decision-fusion strategies, such as weighted soft voting, maintained an accuracy of 96%. Overall, the results demonstrate the feasibility of deep learning for real-time fatigue recognition during table tennis training and provide a practical monitoring tool to help players track their physiological states and reduce injury risk. Full article

Background/Objectives: Dental health is an important component of overall well-being, and accurate tooth analysis supports several applications in dental imaging. Tooth segmentation is a crucial stage in digital dental workflows, particularly for orthodontic diagnosis, treatment planning, and the assessment of treatment-related changes such as root resorption. However, the complex anatomical structures and variations present in dental X-ray images make accurate tooth segmentation and classification challenging. To address these challenges, this study develops a Tooth Segmentation and Classification through the Fusion of Feature Extraction Models and Advanced Optimization Algorithms Using Biomedical Images (TSCFFM-AOABI) framework. Methods: The proposed framework first applies image augmentation using Blur, MedianBlur, ToGray, and Contrast-Limited Adaptive Histogram Equalization (CLAHE) to improve image variability and feature visibility. Tooth instances are subsequently segmented using YOLO11s trained with the Adam optimizer. The segmented tooth regions are then processed through DenseNet201, MobileNet, and ShuffleNetV2 to obtain complementary deep feature representations, which are fused to construct a comprehensive feature representation. The fused features are classified using a Kernel Extreme Learning Machine (KELM), whose penalty parameter C and RBF kernel parameter γ are optimized using Bacterial Colony Optimization (BCO). Results: The proposed framework is evaluated on a benchmark dental X-ray image dataset containing 598 images and 32 tooth classes. Experimental results demonstrate an accuracy of 96.17%, precision of 95.10%, recall of 99.03%, and an F1-score of 98.00%. Conclusions: The results demonstrate the effectiveness of the proposed framework for tooth segmentation and classification. Full article

Artificial Intelligence (AI) is being applied across textile-recycling systems to support material identification, automated sorting, process optimization, recovered-fiber assessment, sustainable composite design, and circular-economy decision-making. This systematic review synthesizes recent developments in AI-driven recycling pathways for sustainable textile composites. A structured Scopus search followed by predefined screening and eligibility assessment yielded 121 studies published between 2018 and 18 June 2026, and the review process was reported in accordance with the PRISMA 2020 framework. The reviewed evidence spans applications ranging from categorical fiber recognition to composition-aware and process-oriented approaches based on machine learning, deep learning, computer vision, spectroscopy, and optimization methods. Near-infrared, hyperspectral, Raman, and image-based systems have shown strong potential for identifying pure and blended textiles, while predictive models have been applied to estimate recovered-fiber quality, optimize recycling conditions, and predict mechanical and functional properties of recycled composites. Mechanical, chemical, and thermal recycling pathways exhibit different trade-offs in feedstock tolerance, material quality, process severity, and recovery potential, indicating that pathway selection remains context dependent. Life-cycle and circularity studies further suggest that higher recovery or recycled content does not necessarily correspond to lower environmental burden, particularly when energy use, substitution potential, material quality, and downstream recyclability are considered. Despite methodological advances, broader industrial implementation remains constrained by data representativeness, ground-truth quality, model transferability, explainability, and end-to-end system integration. Overall, the reviewed evidence suggests that further progress may depend on linking material characterization, quality-aware processing, composite-performance prediction, and sustainability assessment within interoperable and experimentally validated decision-support frameworks. Full article

Data augmentation (DA) is a widely used technique to improve generalization in Deep Learning. However, many existing mixing-based approaches apply uniform strategies or rely on computationally expensive saliency guidance. In this work, we propose AdaptiveIntensityMix (AIM), a data augmentation method that adjusts mixing intensity based on local structural complexity. AIM combines edge magnitude and local variance at multiple scales to identify complex regions, applying conservative mixing to preserve informative structures and stronger mixing to simpler areas. Experiments on six datasets (CIFAR-10, CIFAR-100, Tiny ImageNet, Chest X-Ray, SVHN, and Food-50) show that AIM improves generalization over baseline and performs competitively with Mixup, CutMix, CutOut, and Random Erasing. Comparison with region-aware methods (SaliencyMix, PuzzleMix) shows that AIM achieves comparable accuracy with lower computational overhead. Statistical tests confirm that AIM performs better than baseline (p < 0.05). AIM provides an efficient augmentation strategy applicable across different visual domains. Full article

Ground deformation, especially subsidence, is a geo-environmental hazard that needs to be assessed and predicted in advance to mitigate its effects efficiently and in a timely manner. Interferometric Synthetic Aperture Radar (InSAR) provides high-resolution, all-weather ground deformation monitoring, while advances in Deep Learning (DL) have significantly improved its analytical capabilities. Despite growing interest in integrating DL with InSAR, existing systematic reviews have focused mainly on urban infrastructure, leaving mining-induced subsidence comparatively underexplored. This review addresses this gap by applying the PRISMA framework to systematically evaluate 162 peer-reviewed studies, including 21 core articles on DL-based mining subsidence and 58 supporting studies from related deformation applications. This paper examines DL architectures, including CNNs, LSTMs, GRUs, ConvLSTMs, Transformers, and Graph Neural Networks, and assesses their roles in data decomposition, spatiotemporal feature extraction, hyperparameter optimization, and multi-source data fusion. Comparative analysis of 71 models shows that attention-enhanced CNN-LSTM and ConvLSTM architectures consistently outperform single-model approaches. Key challenges include limited mining datasets, poor cross-site transferability, lack of standardized benchmarks, and insufficient geomechanical integration. Finally, we propose a seven-point research roadmap to support reliable, explainable, transferable, and operational early-warning systems for mining-induced subsidence monitoring. Full article

Background/Objectives: Pulmonary artery (PA) dilatation and slow-flow are imaging features of pulmonary hypertension (PH), but their quantitative assessment on routine cardiovascular magnetic resonance (CMR) remains limited. This study aimed to develop and evaluate an automated CMR pipeline for quantifying PA size-based and slow-flow metrics from routine white-blood (WB) and black-blood (BB) acquisitions. Methods: A total of 205 CMR acquisitions (106 WB and 99 BB) were used for model development and independent testing. Separate deep learning nnU-Net-based segmentation models were developed for WB and BB acquisitions using a two-stage cascaded framework. Automated measurements included PA volume and PA slow-flow. The trained models were subsequently applied to a larger cohort of 1471 patients using available WB and BB acquisitions to assess automated measurement extraction and agreement with clinical reports. Results: Automated PA volumetric measurements demonstrated excellent agreement with manual reference measurements for both the WB and BB models (ICC 0.94 and 0.93, respectively), with high segmentation accuracy (Dice 0.86 and 0.80, respectively). Automated slow-flow measurements also demonstrated excellent volumetric agreement (ICC 0.94), despite lower spatial overlap (Dice 0.61). Automated slow-flow classification demonstrated 83% agreement with qualitative clinical reporting (κ = 0.58). Conclusions: Automated PA volumetric quantification from routine CMR is feasible. Automated slow-flow quantification showed promising performance but requires validation across scanner vendors and BB acquisition protocols before broader clinical application. Full article

The rapid advancement of deep learning has significantly improved the performance of remote sensing object detection. However, detection performance for small objects remains unsatisfactory. In this paper, we identify frequency aliasing as a key challenge underlying small object detection performance degradation and propose DWT-DETR, a novel frequency decoupling and recalibration small object detector using discrete wavelet transform (DWT). The proposed DWT-DETR employs DWT to obtain explicit frequency-band representations and further models them according to the different requirements of feature encoding and multi-scale feature fusion. In the feature encoding stage, we design a frequency-band mixing module (DWT-FBM) to mitigate the degradation of high-frequency details of small objects in deeper layers. The DWT-FBM first employs Haar wavelets to decompose features into frequency sub-bands, then applies cross-frequency interaction and dynamic amplitude modulation to adaptively recalibrate the frequency information. Finally, the refined features are reconstructed via inverse wavelet transform. During feature fusion, we design a frequency-guided fusion module (DWT-FGF) to ensure spatial alignment and discrimination of the fused features. The DWT-FGF employs a dual-path mechanism that decouples shallow features into low-frequency and high-frequency sub-bands via wavelet decomposition. The low-frequency sub-band serves as a spatial reference to guide deep upsampling and reduce geometric deformation, while the high-frequency sub-bands are used to suppress background noise. Extensive experiments on AI-TOD, DIOR, and VisDrone datasets demonstrate the effectiveness of DWT-DETR, which achieves 30.8% AP, 79.4% mAP, and 32.6% AP, respectively. Full article

This bibliometric study examines research that explicitly combines sparse methods with Tucker-related techniques. Using PRISMA 2020 as a reporting guide for record identification and selection, we retained 254 journal articles indexed in Scopus and Web of Science and published between 2000 and May 2025. The analysis covers annual publication output, raw citation counts, international collaboration, and keyword co-occurrence using VOSviewer (version 1.6.20) and R (version 4.4.2). The 254 articles show sustained growth in publication output between 2010 and 2024. China and the United States account for a substantial share of the retrieved literature, while Pattern Recognition and IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing are prominent within this specific dataset. Because citation counts vary with publication year, discipline, and database coverage, they are interpreted as descriptive indicators rather than direct measures of research quality or normalized impact. The retrieved studies include applications in neuroscience, finance, hyperspectral image analysis, and multimodal learning. The co-occurrence maps center on terms such as tensor decomposition, low-rank approximation, and deep learning. These results describe how sparse and Tucker techniques are being used across several areas of data science, while remaining specific to the corpus defined by the search strategy. Full article

Identifying the specific residues that initiate protein folding is critical for understanding protein stability and the mechanisms of misfolding-related diseases. However, the systematic, proteome-wide identification of these early folding residues (EFRs) has been hampered by both experimental and computational limitations. Here, we present LMEFold, a deep learning framework using protein language models (PLMs) to accurately predict EFRs from amino acid sequence alone. LMEFold outperforms existing sequence-based predictors on the Start2Fold benchmark (ROC AUC = 0.831), and its predicted scores show consistent associations with independent HDX-NMR protection-factor measurements, providing complementary biophysical evidence. Applying LMEFold to annotate over 6.6 million variants across diverse genomic datasets, we find that pathogenic mutations curated in ClinVar are significantly enriched within EFRs (~17%) compared to millions of population variants from large-scale cohorts including the UKB and gnomAD databases (~10%, P < 0.001). Critically, in the MSK-MET pan-cancer clinical cohort, EFR mutations are associated with severe functional disruption and significantly reduced overall patient survival. Structural analysis of hotspot cancer mutations, such as TP53 p.Y163C, suggests a mechanism whereby these variants destabilize protein structure by disrupting the early folding core. Our work establishes EFRs as critical sites of pathogenic vulnerability and provides a powerful framework for the rapid identification of folding-centric prognostic markers using sequence only. Full article

Deep learning has been extensively applied to the modeling and management of urban flooding. While deep learning provides the analytical framework, the unique characteristics of the underlying data modalities themselves fundamentally drive model performance. Thus, this paper provides a scoping review of 289 recent studies (2016–2025), categorizing the relevant work into five modalities based on data acquisition methods, characteristics, and information processing workflows: Hydro-Geospatial Physical Base, Remote Sensing Imagery, Human Perception and Social Sensing, Environmental Acoustic, and Cross-Modal Integration. The hydro-geospatial physical base and remote sensing imagery dominate current studies because they are stable, widely available, and information-rich, whereas human behavior and acoustic data remain underutilized owing to heterogeneity, limited availability, and costly preprocessing. Key bottlenecks include information loss during preprocessing, insufficient cross-modal alignment, representations that struggle to balance statistical correlation with physical consistency, and marked regional variations in data. Ultimately, these limitations indicate that the field is still primarily dominated by purely data-driven models, and future progress requires integrating statistical learning with physical mechanisms to improve robustness and generalization capabilities. Full article

Background: Composite materials are not easy to model using traditional experiments and physics-based approaches due to their diverse properties and multiscale nature. This review summarizes how artificial intelligence (AI) methodologies, including machine learning, deep learning, and generative AI, have been increasingly transforming and how they have been increasingly applied to composite materials research. It combines their possible use across the prediction of their properties, damage detection, structural health monitoring, and materials design. Methods: Searches were performed in Scopus and Google Scholar for articles published using combined keywords such as “(artificial intelligence OR machine learning OR deep learning OR Generative AI) AND (composite OR fiber reinforced) AND (prediction OR optimization OR characterization)”. The search strategy was designed to capture both fundamental developments in AI methodologies relevant to materials research and their specific applications to composite materials. The identified publications underwent a two-step selection process: (i) an initial review of titles and abstracts, and (ii) a detailed full-text evaluation. Studies were included if they provided explicit descriptions of AI models, clearly specified datasets, and applied quantitative performance assessments to composite materials-related applications such as material property prediction, damage detection, characterization, or manufacturing process optimization. For each selected article, relevant information was extracted in a structured and consistent manner, including the type of composite material, AI methodology, input parameters, dataset characteristics and size, validation methods, and application domain. These data were used to support both quantitative trend analysis and the qualitative identification of research gaps and future research opportunities in composite materials research. Results: Supervised learning methods, in particular artificial neural networks, support vector machines, random forests, and decision trees, were commonly used and exhibited durable potential accuracy in predicting properties of composite materials. Unsupervised methods such as principal component analysis (PCA) and clustering were comparatively underused. Generative AI, including generative adversarial networks (GANs) and variational autoencoders (VAEs), and physics-informed models emerged as growing approaches for synthetic data generation and improved interpretability, respectively. Limitations: The reviewed studies were frequently limited by small or heterogeneous datasets, weak generalizability testing beyond training conditions, high computational requirements and limited model interpretability (“black box” behavior); a risk-of-bias assessment across included studies was not formally conducted. Conclusions: AI is shifting composite materials research from trial-and-error toward data-driven, predictive design. Realizing its full potential will require open, well-annotated datasets, physics-aware and explainable models, and closed-loop, active-learning workflows linking prediction to experimental validation. Full article

We present GalSpecCNN, a probabilistic deep-learning framework for estimating fiber-scale stellar mass ($M_{*}$) and specific star formation rate (sSFR) from LAMOST galaxy spectra. The model uses a one-dimensional convolutional neural network to extract spectral features, followed by a probabilistic regression head that predicts a diagonal bivariate Gaussian distribution. Aleatoric uncertainty is modeled through the predicted variances, while epistemic uncertainty is approximated using Monte Carlo Dropout. On the test set, GalSpecCNN achieves residual scatters of 0.226 dex for $M_{*}$ and 0.319 dex for sSFR, with small mean residuals. The empirical coverages of the nominal 68% predictive intervals are 68.60% for $M_{*}$ and 71.58% for sSFR, indicating good overall calibration with mildly conservative coverage for sSFR. We apply the trained model to approximately $1.16\times {10}^5$ LAMOST DR7 galaxy spectra and construct a catalog of $M_{*}$ and sSFR estimates, together with their associated predictive intervals. This work provides a practical framework for uncertainty-aware estimation of galaxy physical properties from large spectroscopic surveys. Full article

Objectives: This study aimed to develop and validate a deep learning–based system for automated detection of intervortex venous anastomosis (IVA) on ultra-widefield indocyanine green angiography (UWF-ICGA). Methods: A total of 183 UWF-ICGA images with an original resolution of 3900 × 3072 pixels were included. A two-stage deep learning framework was developed. In the first stage, an object detection model automatically localized four predefined quadrants within each UWF-ICGA image. In the second stage, each extracted quadrant was classified according to the presence or absence of IVA. Quadrants demonstrating definite venous connections between adjacent vortex vein drainage territories were defined as IVA-positive, whereas quadrants without identifiable intervortex venous connections were defined as IVA-negative. Brightness normalization was applied before analysis, and images were resized to 512 × 512 pixels using linear interpolation. The object detection model was trained using 155 images and tested using 28 images. The 183 original images generated 732 quadrant images, of which 622 were used for training and 110 for testing of the classification model. Model performance was evaluated using accuracy, precision, recall, F1 score, and confusion matrices. Results: The object detection model correctly localized all predefined quadrants in the test dataset, achieving an accuracy, precision, recall, and F1 score of 100.0%. All 112 manually labeled regions from 28 test images were correctly matched by the model. The IVA classification model achieved an accuracy of 93.63%, precision of 95.39%, recall of 91.46%, and F1 score of 93.38%. Among 110 test quadrants, all 69 IVA-negative quadrants were correctly classified, while 34 of 41 IVA-positive quadrants were correctly identified. Seven IVA-positive quadrants were misclassified as negative, whereas no IVA-negative quadrant was classified as positive. When IVA-positive was considered the positive class, the corresponding sensitivity and specificity were 82.9% and 100%, respectively. Conclusions: A two-stage deep learning system enabled automated detection of IVA on UWF-ICGA with high classification performance. Automated assessment of IVA may provide an objective and reproducible approach for evaluating choroidal venous remodeling and may serve as a platform for future quantitative investigations of vortex vein abnormalities. Full article

Achieving high-resolution imaging for a deployable space telescope requires segments co-phased to nanometer precision. Typical alignment uses point sources (distant stars) or extended ground scenes. An on-board active correction system has been developed to provide continuous alignment by using a fibre source instead. IMPACT (Image-based Mirror Phasing and Alignment using Convolutional neTworks) measures point spread functions (PSFs) on a secondary mirror detector separated from the main science camera. By applying a deep machine learning algorithm, piston and tip/tilt aberrations can be retrieved and corrected for. When simulated from a uniform distribution, PSF images were corrected from a mean Strehl of $0.18$ to $0.99$ after two model passes, reducing RMS errors from a mean of $276.7$ nm down to $11.1$ nm. Full article

Consumer electronics powered by on-device artificial intelligence face a structurally coupled dual threat: adversarial attacks that manipulate model outputs, and side-channel attacks that exploit physical implementation characteristics to extract cryptographic keys and neural network weights. Prior work addresses these threats independently, leaving exploitable gaps when both attack vectors are applied in a coordinated manner. This paper presents ChipGuard-AI, a hardware-assisted unified security framework that co-designs adversarial input detection, multi-modal side-channel countermeasures, and Physical Unclonable Function-rooted device authentication on a shared cross-domain security substrate. A dedicated Security Processing Unit maintains a cross-domain security state vector in real time, enabling coordinated hardware-accelerated response to concurrent multi-vector attacks with 2.3 ms adversarial detection latency. A Deep Q-Network adaptive policy dynamically allocates countermeasures to balance security strength against resource overhead. Experimental validation on CIFAR-10 adversarial benchmarks and ChipWhisperer physical side-channel traces demonstrates 97.3% adversarial detection accuracy and 89.7% side-channel leakage reduction. Field-programmable gate array synthesis on Xilinx Zynq UltraScale+ achieves these security levels at 12.4% computational overhead and 6.8% power increase. The adaptive reinforcement learning policy achieves 11.3% lower average overhead than threshold heuristics while maintaining equivalent security effectiveness. Multi-platform validation across four consumer electronics device classes, from ultra-low-power Internet of Things endpoints to flagship smartphone system-on-chips, demonstrates 91.2% to 95.7% adversarial detection and 76.4% to 85.3% side-channel protection with 8.3% to 14.1% computational overhead. Full article