Search Results (4,774)

Accurate and timely identification of slippery road surfaces is essential for ensuring driving safety and operational efficiency on highways. However, blurred vehicle-background boundaries, uneven illumination, and water splashing caused by passing vehicles make existing image-based recognition methods prone to low accuracy. To address these challenges, this paper proposes an improved SegFormer-based framework with two task-specific innovations: (1) a novel Guided Multi-scale Fusion (GMF) module for task-guided multi-scale feature integration, designed to incorporate auxiliary information such as vehicle type, relative speed, and splash regions, enabling the network to focus on slipperiness-relevant road areas while suppressing background interference; and (2) an improved Boundary Attention Awareness (BAA) module with directional Sobel-based boundary initialization, which provides explicit geometric priors to preserve fine boundary details and reduce ambiguity in slippery regions with irregular or weak edges. A multi-scale input and enhancement strategy is further adopted, along with a weighted combination of cross-entropy loss and Dice loss to mitigate class imbalance. Experimental results on our self-constructed Guangzhou Beierhuan Expressway dataset achieve an mIoU of 95.80%, accuracy of 97.84%, and F1-score of 97.86%. To verify cross-domain generalization, we further evaluate the model on two additional benchmarks: it achieves an mIoU of 93.51% on the synthetic SYN-UDTIRI dataset, and attains an mIoU of 95.80% with an AmIoU of 76.20% on the public Cityscapes dataset, achieving competitive performance against several mainstream architectures. The proposed method offers considerable application potential for highway safety warning systems. Full article

Electroencephalography (EEG) provides a non-invasive means of supporting Attention-Deficit/Hyperactivity Disorder (ADHD) assessment. Nonetheless, existing pipelines often rely on handcrafted descriptors, segment-wise decisions, or deep architectures with limited subject-level generalization. This work introduces Hidden Markov Model-Induced Stationary RKHS Distance Learning (HIS), a probabilistic framework that represents each subject by a Hidden Markov Model with Gaussian-mixture emissions trained directly from frontal EEG recordings. Rather than vectorizing model parameters, each HMM is mapped to its induced stationary observation distribution and embedded into a Reproducing Kernel Hilbert Space (RKHS), where pairwise subject similarities are computed through a closed-form Hilbert embedding distance. These similarities are subsequently exploited by precomputed-kernel classifiers for subject-level prediction. The proposed method was evaluated against the Probability Product Kernel baseline using both a controlled synthetic EEG benchmark and a public pediatric ADHD dataset under progressively more rigorous validation protocols, culminating in repeated nested cross-validation with bootstrap confidence intervals and permutation testing. On the synthetic benchmark, HIS achieved 95.0% held-out accuracy and consistently outperformed the baseline across classifiers. On a real EEG dataset with 121 subjects, the primary evaluation protocol yielded a balanced accuracy of 73.5% (95% CI: 69.8–77.0%), an AUC of 79.6%, and an MCC of 0.483 (permutation p < 0.001) using an SVM with compact subject-specific HMMs. Complementary hyperparameter analyses and t-SNE visualizations demonstrated that HIS induces more stable and discriminative subject representations than the baseline. These results establish stationary RKHS embeddings of subject-specific HMMs as a leakage-aware framework for EEG-based ADHD decision support and underscore the critical influence of statistically rigorous evaluation protocols on reported classification performance. Full article

Segmenting key components of space targets using Inverse Synthetic Aperture Radar (ISAR) images is an important interpretation task in space situational awareness. However, the scarcity of pixel-level annotated data, inter-class confusion caused by morphological differences among multiple target classes, and the absence of structural priors for components restrict the performance improvement in existing deep models on this task. Therefore, this paper proposes a Semi-Supervised Structural Prior-Guided Network (SSPNet). First, a Gated Manifold-Constrained Hyper-Connections Vision Transformer (GMHC-ViT) encoder is proposed to broaden the feature representation space via parallel multi-feature streams with adaptive gating, thereby alleviating inter-class confusion and enhancing cross-category generalization. Second, a Prior-Guided Module (PGM) is proposed to extract shape and edge priors of components, and it adaptively enhances the weakly activated channels of encoder features through cross-attention, thereby injecting structural knowledge independent of image quality into the segmentation process. Furthermore, to effectively leverage large amounts of unlabeled data, a strong perturbation strategy tailored to the characteristics of ISAR images is designed for consistency regularization. Experimental results on a simulated ISAR dataset containing 38 classes of space targets demonstrate that SSPNet outperforms existing methods and exhibits strong segmentation capability even under low signal-to-noise ratio (SNR) conditions. Full article

Soil pesticide-residue screening is important for ecological protection, food safety, and public health. However, conventional chromatographic and spectroscopic methods often require complex sample pretreatment, expensive instruments, trained operators, and long analysis times, which limits their use for rapid and large-scale screening. In this study, an electronic-nose system was developed for volatile-fingerprint classification of pesticide-treated loess soil. Six commercial pesticide formulations from three chemical categories were evaluated: deltamethrin and cyfluthrin as pyrethroids, glyphosate and chlorpyrifos as organophosphorus pesticides, and zineb and mancozeb as organosulfur pesticides. These compounds were selected to represent commonly used pesticides with different chemical structures and volatile profiles. A mirror-symmetric gas-sensing chamber was designed for a 26-sensor metal oxide semiconductor (MOS) array to improve gas-flow uniformity and response repeatability. A total of 960 pesticide-treated electronic-nose response curves were collected from four soil depths. Eight feature extraction methods and four classifiers were compared. The Synthetic Minority Over-sampling Technique (SMOTE) and Geometric SMOTE (G-SMOTE) were then evaluated using training-fold-only oversampling to reduce data-leakage risk in imbalanced classification. The results showed that k-nearest neighbors (KNN) combined with direct or transform-based features provided strong classification performance under controlled laboratory conditions. For minority-class recognition, SMOTE showed more consistent improvement than G-SMOTE in the tested pesticide–depth–feature combinations, although the effect depended on feature representation and pesticide class. These findings indicate that the proposed chamber/sensor-array/SMOTE framework is feasible for rapid volatile-fingerprint classification of pesticide-treated soil, but larger independent field datasets and quantitative chemical validation are still required before general deployment. Full article

This study presents a modular AI-assisted workflow for converting single 2D interior images into textured 3D assets and for evaluating those assets when ground-truth 3D meshes are unavailable. The proposed pipeline combines object detection, instance isolation, monocular-depth estimation, image-to-3D generation, texture synthesis, mesh export, and cloud-based execution to support early-stage interior-design and real-estate visualization tasks. A reference-free validation protocol is introduced, based on rendered multi-view comparisons, silhouette Intersection-over-Union, automated captioning, and multimodal embedding similarity, and is complemented by a composite validation framework that benchmarks reconstructed scenes against 200 panoramic indoor scenes from the Structured3D dataset using Hungarian-matched placement, size, recall, and relative-distance metrics. The workflow was implemented and tested using contemporary computer-vision and generative 3D components, with Hunyuan3D 2.0 used as the main reconstruction model. Proof-of-concept experiments on a representative corpus of 178 synthetically generated single-object images spanning a range of interior furniture categories show comparable silhouette IoU for textured and non-textured outputs and indicate that texture-preserving renderings improve visual and semantic similarity scores across CLIP-based evaluations. The 200-scene dataset evaluation reveals stable spatial localization (placement error ≈ 1.18 m, relative-distance error ≈ 0.54 m) alongside systematic over-prediction and size-calibration errors. Beyond the applied pipeline, the study contributes a reference-free, ground-truth-free protocol for 3D-asset evaluation and a first quantified account of where object-centric single-image reconstruction is reliable—spatial placement—and where it is not—object scale and spurious detection—at interior-scene scale. The results demonstrate the feasibility of integrating perception, 3D reconstruction, semantic assessment, and scalable deployment into a single applied pipeline, while remaining proof-of-concept and requiring extension to larger object and scene corpora, baselines, real-photograph evaluation, and human-centered assessment before broad claims about general interior-scene reconstruction can be made. Full article

Power quality disturbance (PQD) classifiers can lose 30–50 percentage points of accuracy when the signal-to-noise ratio (SNR) drops to 5 dB. This study finds that the degradation arises primarily not from architectural limitations but from hyperparameter configurations that fail to adapt across noise levels, within the IEEE 1159 disturbance set and additive-noise conditions studied here. Building on this insight, we propose a one-dimensional convolutional neural network (1D-CNN) with channel attention, whose three key hyperparameters—the learning rate, the first-layer convolutional kernel size, and the attention reduction ratio—are automatically optimized using the Phong Optimization Algorithm (POA). A 12-class synthetic dataset constructed to IEEE 1159-2019, with noise levels from noise-free to 5 dB SNR, is used for training and evaluation. Under the extreme condition of 5 dB SNR, the proposed method achieves 83.4% accuracy, outperforming a support vector machine (SVM) with discrete wavelet transform features (66.30%), a hybrid CNN–long short-term memory network (CNN-LSTM; 43.32%), and plain 1D-CNN (38.39%) under their commonly reported configurations. When every deep-learning baseline receives the same POA hyperparameter optimization under a fair per-SNR protocol, this advantage largely disappears: at 5 dB SNR all POA-optimized deep methods fall within roughly 5 percentage points (82.31–87.13%), and the attention module’s own contribution shrinks to within run-to-run variation (the same proposed model scoring 82.77% with attention vs. 82.48% without), showing that systematic hyperparameter optimization, not architectural novelty, drives the noise robustness. A single-set ablation under the original 10 dB-optimized configuration points the same way (POA optimization alone raising 5 dB accuracy from 41.6% to 76.4%, with channel attention adding a further 7 points). A comparison against classical threshold-index classifiers on the identical dataset shows the same pattern at the level of hand-crafted features: with fixed clean-calibrated thresholds, the index classifier collapses from 95.7% to 17.5% at 5 dB SNR, while per-noise-level re-calibration of the same indices recovers 83.8%. Furthermore, the POA-optimized hyperparameters were validated on an independent public PQD dataset, achieving 90.00% accuracy when training a fresh model from scratch. It also stays robust under more realistic complex noise (80.9% at 5 dB SNR) and, on a two-class real-measured probe, transfers to field signals with only light calibration. These findings suggest that for noise-robust PQD classification, hyperparameter optimization deserves as much attention as the architectural design itself, rather than being treated as a final tuning step. Full article

This paper presents a novel approach for estimating 3D human pose from monocular video sequences in underwater scenarios, tackling the unique challenges posed by water refraction, body occlusion, low image quality and illumination distortion in underwater environments. Leveraging both 2D keypoint extraction and parametric model estimation, our method operates in a two-stage framework including preprocessing and optimization. In the preprocessing stage, a Part Attention Regressor (PARE) is adopted to dynamically estimate SMPL human body parameters, particularly adept at handling occlusions common in underwater scenarios. Additionally, a 2D keypoint detector, employing YOLO for bounding box detection and HRNet for keypoint regression, enhances feature extraction despite underwater image challenges. In the optimization stage, we propose an underwater variational autoencoder (UW-VAE), which adopts a data-driven strategy to learn the biomechanical prior distribution of underwater human poses and implicitly correct unreasonable pose parameters caused by refraction and occlusion. The optimization process incorporates constraints aligning final SMPL models with detected 2D keypoints, minimizing disparity between adjusted and original SMPL models, and ensuring temporal consistency. Furthermore, to address the scarcity of annotated underwater datasets, we build a full pipeline to generate synthetic underwater datasets with complete annotations based on UW-VAE. Experimental results on the SwimXYZ synthetic dataset show that our method achieves 51.60% PCK@0.2 and 80.13% PCK@0.5, outperforming state-of-the-art land-based methods across most stroke categories. Validation on real-world underwater swimming datasets demonstrates improved 2D keypoint accuracy after synthetic-data fine-tuning, which provides a new solution for 3D human motion analysis in underwater sports, biomechanical research and swimming training. Full article

Three-dimensional synthetic aperture radar (SAR) point clouds provide valuable geometric observations of urban scenes, but they often suffer from severe noise and layer-like artifacts caused by the low signal-to-noise ratio and tomographic imaging mechanism. These degradations make SAR point cloud denoising significantly more challenging than conventional LiDAR point cloud denoising. In this paper, we propose a Geometry-guided Diffusion SAR Point Cloud Denoising (GDSD) framework to recover geometrically coherent building surfaces from noisy SAR point clouds.The key idea is to exploit relatively clean LiDAR point clouds as geometry priors while avoiding the need for paired SAR–LiDAR supervision or clean SAR ground truth. Specifically, we introduce a Forward Gaussian Noising Process to disrupt the intrinsic layer-like artifacts of SAR point clouds and reduce the input-level discrepancy between SAR and LiDAR domains. We further design a geometry prototype-based alignment module that projects SAR and LiDAR bottleneck features into a shared LiDAR-dominated latent space, enabling geometry-aware conditional reverse diffusion. A DiT-3D-based denoising network is then trained with LiDAR-domain diffusion supervision and applied to SAR point clouds using the aligned SAR geometry condition. To evaluate the proposed method, we construct a SAR point cloud denoising benchmark based on the MV3DSAR dataset with CAD-derived reference surfaces. Experimental results show that GDSD significantly improves the quality of noisy SAR point clouds and clearly outperforms the previous conventional LiDAR point cloud denoising baseline, producing more continuous and geometrically coherent SAR building point clouds. Full article

This study evaluated oxidative damage, antioxidant-related, and nitric oxide (NO)-related chemical markers in commercial bee pollen using two extraction systems. For this purpose, a total of 30 bee pollen samples from different brands and/or manufacturers were examined; each sample was subjected to aqueous extraction and ethanol/water (80:20, v/v). The levels of reduced glutathione (GSH), malondialdehyde (MDA), and NO metabolites, comprising nitrite and nitrate and expressed as total NOx, were determined in pollen extracts using spectrophotometric methods. The data were analyzed using descriptive statistics, group comparisons, and Spearman correlation analysis. The findings showed that ethanol/water (80:20, v/v) extraction provided significantly higher measured GSH equivalents and total NOx levels than aqueous extraction. Mean GSH equivalents were 79.12 ± 25.54 µmol/g in the aqueous extracts and 189.77 ± 45.09 µmol/g in the ethanol/water extracts, and the difference was found to be statistically significant (p < 0.001). NO metabolite levels were determined as 5.97 ± 8.34 and 366.70 ± 85.10 µmol/g, respectively, and the difference between the extraction methods was found to be statistically significant (p < 0.001). In contrast, MDA levels did not show a significant difference between the two extraction systems (p = 0.391). A significant positive correlation was observed between GSH and NO metabolites in the combined dataset (r = 0.660, p < 0.001). The findings suggest that 80% ethanolic bee pollen extract may represent a promising source of NO metabolites for future investigation as a natural alternative to synthetic sodium nitrite in processed meat products. However, this hypothesis requires validation through microbiological, technological, sensory, and shelf-life studies. Full article

The rapid proliferation of Internet of Things (IoT) devices has accelerated the development of Smart Home environments. However, research in this domain remains constrained by the scarcity of high-quality datasets and the challenges associated with collecting real-world data. The HESTIA simulator addresses this issue by generating realistic synthetic data, but its adoption is hindered by a complex configuration process that requires manual editing and technical expertise. To address these limitations, this work introduces HestiaInteract, a web-based interaction layer that abstracts the low-level configuration required by HESTIA. Rather than merely providing a graphical interface, the proposed platform introduces a structured workflow for semantic scenario modeling, reusable simulation components, automatic input validation, and browser-based execution. By replacing manual JSON editing and script-based configuration with guided visual interactions, the platform reduces technical barriers while preserving the flexibility of the original simulator. A usability study involving 50 participants was conducted to evaluate the proposed solution. The results indicated high levels of satisfaction, ease of use, interaction quality, and clarity in task execution. Participants reported that the interface significantly simplified the simulation process and improved the overall user experience. These findings demonstrate that HestiaInteract enhances the usability and accessibility of the HESTIA simulator, facilitating its adoption and supporting research in intelligent environments. Full article

Background: Increasing access to Electronic Health Records (EHRs) has enabled the development of Machine Learning (ML) models to predict early cardiovascular disease (CVD) risk. Nevertheless, real EHR data is usually limited in availability and distribution because of confidentiality issues, legal limitations, and insufficient access. Synthetic data generation has become a promising approach to overcome such difficulties. Objectives: This study proposes a Large Language Model (LLM)-guided retrieval-aware framework for generating realistic synthetic EHR data on a large scale to train ML models to predict cardiovascular risks accurately. Methods: The framework uses a Tabular Denoising Diffusion Probabilistic Model to learn the underlying distribution of the original dataset and generate an initial synthetic dataset. To improve the clinical plausibility of the generated data, a knowledge-guided refinement module with LLaMA 2 13B combined with Retrieval-Augmented Generation (RAG) is introduced. The LLM analyzes statistical trends in real and synthetic data while retrieving relevant medical information to detect and correct clinically implausible correlations. Results: Experimental findings show that the optimized synthetic dataset preserves important statistical features of the original data and achieves high performance in predicting CVD. Conclusions: Thus, the framework offers a scalable, privacy-conservative method for generating realistic synthetic healthcare data suitable for medical research. Full article

Limited sample diversity poses a major challenge for deep learning-based coal–gangue recognition using mineral laser speckle images under complex industrial conditions. To address this issue, we propose an improved data augmentation framework based on a Deep Convolutional Generative Adversarial Network (DCGAN). Dense Blocks are introduced to enhance high-dimensional feature representation without significantly increasing network complexity, while self-attention modules strengthen multiscale feature interactions between localized flare features and global speckle distributions. Residual-based sampling modules and the Wasserstein GAN with Gradient Penalty (WGAN-GP) objective are further employed to improve gradient propagation and adversarial training stability. Experimental results on synthesizing 512 × 512 mineral laser speckle images show that the proposed model reduces the Fréchet Inception Distance (FID) by 72.75% relative to the baseline DCGAN. Furthermore, the YOLOv5 detectors trained on the augmented datasets achieve a peak mean Average Precision (mAP) of 99.3%, significantly outperforming conventional geometric transformation methods and typical generative baselines including StyleGAN2-ADA and SAGAN. These results demonstrate that the proposed method effectively improves alignment between the synthetic and real distributions, enhances speckle feature diversity and boosts the accuracy and robustness of coal–gangue recognition under limited-sample conditions. Full article

Background and Objective: Maximal oxygen uptake (VO₂max) is a fundamental indicator of cardiorespiratory fitness in sports medicine, essential for athletic profiling and training prescription. However, traditional direct measurements require exhaustive physical testing, which induces considerable physiological stress, limits testing frequency, and increases the risk of injury. To address this problem, this study aimed to eliminate the need for exhaustive protocols by developing a highly precise, non-invasive digital prediction model for VO₂max utilizing readily available anthropometric data and acute metabolic biomarkers (blood glucose and lactate kinetics). Methods: To overcome the limitations of a small initial empirical sample (n = 16) and prevent model overfitting, the original dataset was statistically augmented to create a robust synthetic cohort (n = 200) using Multivariate Normal Distribution and k-Nearest Neighbors (k-NN) algorithms. Three different machine learning models (Multiple Linear Regression [MLR], Random Forest [RF], and Support Vector Regression [SVR]) were trained using such parameters as sex, height, weight, baseline/pre-exercise, and net (Δ) glucose and lactate concentrations. Evaluation of the performance of the models included tenfold cross-validation and Bland–Altman analysis as a measure of clinical agreement. Results: Among the three algorithms used, the highest correlation coefficient (R² = 0.939) was observed for SVR, along with the lowest error metrics (RMSE = 1.442 mL/kg/min, MAPE = 2.78%). Moreover, SVR showed remarkable performance in predicting VO₂max of female (R² = 0.890) and male (R² = 0.702) athletes separately. Also, Bland–Altman analysis proved almost zero-bias estimation with 95% limits of agreement ranging between −4.40 and 4.43 mL/kg/min. In order to bypass the black-box problem of complex algorithms for practical application in the field, the MLR model was used for the development of the Beta-Weighted Lacto-Glycemic Equation (β-LGE). Conclusions: In this work, a unique dual-model approach was introduced, with the SVR algorithm serving as a high-performance backend of a digital tool for sport technologists and Beta-Weighted LGE providing a practical calculation formula for coaches. This innovative approach allowed for a completely non-exhaustive profiling of an athlete’s VO₂max using only minimally invasive metabolic measurements. Full article

This study includes an interpretable machine learning (ML) framework for predicting the shear contribution of externally bonded fiber-reinforced polymer (FRP) composites in reinforced concrete beams. A database including total 313 experimental specimens was collected from previous experimental research. The data screening process has been conducted using the Isolation Forest algorithm, resulting in 268 cleaned specimens. The cleaned database was divided into a training subset containing 214 specimens and an independent test set containing 54 specimens. The trained subset was enlarged into 5204 synthetic data using two advanced generative models including Wasserstein generative adversarial network and conditional Variational autoencoder (CVAE). Separate constrained monotonic neural network (CMNN) models were then trained on both datasets and WGAN-based CMNN achieved R² = 0.9524 or the synthetic training dataset and R² = 0.9120 for the independent test set, whereas the CVAE-based CMNN achieved corresponding values of 0.9632 and 0.9011. To improve practical applicability, response functions were extracted from WGAN-based CMNN and fitted with analytical expressions to derive a closed-form prediction equation. The proposed equation was independently validated using separate unseen test specimens, which were not used in CMNN training and achieved R² = 0.79, RMSE = 24.98 kN, MAE = 19.65 kN, MAPE = 21.72%, VAF = 79.35%, U95 = ±54.94 kN, SI = 3.04, and PI = 0.11. Compared with ACI 440.2R-17, CSA-S806.12, CNR-DT200 R1.2013, TR-55, and JSCE, the proposed equation showed superior accuracy while maintaining a transparent and design-oriented format. Full article

In ground-based space surveillance, space target attitude estimation is critical for space situational awareness, yet existing methods based on inverse synthetic aperture radar (ISAR) images suffer from three core limitations: phase information is discarded in amplitude-only processing, convolutional neural networks have a restricted global receptive field, and the physical topology of satellite components is not explicitly modeled. To address these issues, we propose a complex-domain Transformer–graph neural network (CD-TrGNN) that unifies global context modeling and adaptive topological reasoning in an end-to-end framework. Specifically, a complex-domain Transformer module (CD-Transformer) with tailored attention captures long-range dependencies among image patches while preserving both amplitude and phase information; a complex-domain graph convolution module (CD-GC) with learnable adjacency matrices and a dual-path update mechanism explicitly encodes the structural relationships among satellite parts. On a self-built ISAR complex image dataset, CD-TrGNN achieves a three-axis mean absolute error of only 1.70°, substantially outperforming six representative baselines. Ablation experiments confirm the effectiveness of complex-domain processing, global attention, and topological reasoning. At a 5 dB signal-to-noise ratio, the error remains at 2.81°, and the accuracy stays below 2° for two different satellite structures. These results demonstrate that CD-TrGNN can fully exploit the information in ISAR complex images, enabling high-accuracy and highly robust attitude estimation. Full article