Search Results (7,102)

Reliable hourly forecasts of fine particulate matter (PM_2.5) can help cities anticipate deteriorating air quality and prepare measures to limit exposure. A central modeling challenge is to combine rapid pollution fluctuations with meteorological and ecological observations that are updated less frequently. This study develops an adaptive frequency gated network (AFGN) that integrates hourly pollution histories, daily meteorological records, and satellite ecological indicators. Parallel temporal convolutions and recurrent encoders represent concentration changes at multiple resolutions, while attention and signed gating incorporate environmental conditions into the forecast. The framework is evaluated using observations from Beijing, Shanghai, Chengdu, and Wuhan during 2023 and 2024. AFGN achieves lower mean prediction error than the gated recurrent unit (GRU) baseline at 1, 6, and 12 h. At 6 h, mean root mean square error (RMSE) decreases from 13.498 to 12.526 μg/m³, corresponding to an absolute reduction of 0.971 μg/m³ and a relative reduction of 7.20%. These results demonstrate the value of combining multiresolution pollution representations with environmental context for short-range forecasting. The approach provides concentration estimates that could inform the timing of pollution surveillance and exposure advisories as part of sustainable urban air quality management. Full article

This study investigates the application of Ka-band (35 GHz) microwave non-destructive testing (NDT) to detect and classify surface-breaking defects in metallic pipelines and both internal and external defects in non-metallic pipelines. A microwave scanning system utilizing an open-ended rectangular waveguide sensor collects 420 one-dimensional voltage line-scan measurements across six defect types (hole, cut, corrosion, internal wear, external wear, and inclusion) in PVC, HDPE, and metallic pipe specimens at three standoff distances (1 mm, 2 mm, and 3 mm). A one-dimensional convolutional neural network (1D CNN), trained with SMOTE resampling and $4\times$ data augmentation, is evaluated under a grouped protocol in which all six scans of any physical defect specimen are confined to a single partition. The network takes the scan through a convolutional branch and receives the standoff distance and material code as separate auxiliary inputs. Grouped five-fold cross-validation, repeated five times over independent partitions, gives a mean accuracy of 82.19 ± 5.93% with a macro-averaged F1-score of 0.85. Errors are concentrated among hole, cut, and external wear, while corrosion, inclusion, and internal wear are recovered with recall at or above 0.90. Validation used artificially introduced, isolated defects in laboratory specimens under controlled acquisition conditions. These results establish laboratory feasibility for Ka-band microwave NDT combined with a 1D CNN as a noninvasive route for automated pipeline defect classification, and show that a learned classifier resolves signal ambiguity that makes manual interpretation of amplitude-only microwave sensor outputs difficult. Full article

Brain tumor classification from magnetic resonance imaging (MRI) can support computer-aided diagnosis by enabling automated differentiation of tumor and non-tumor categories. This study presents DualXBrNet, a hybrid dual-level fusion framework for multiclass brain tumor classification using the original BRISC2025 dataset. Three pretrained convolutional neural networks (CNNs), ResNet50V2, MobileNetV2, and EfficientNetB0, were optimized to learn complementary representations of glioma, meningioma, pituitary tumor, and no-tumor MRI images. Their deep features were integrated through a learnable Feature Fusion network, while the class probability outputs of the three CNNs and the Feature Fusion model were subsequently combined using a leakage-aware five-fold out-of-fold (OOF) decision-level stacking strategy. A low-complexity Logistic Regression meta-classifier was trained exclusively on the resulting 16-dimensional OOF probability representations to prevent information leakage during meta-learning. On the independent official test set, DualXBrNet achieved 96.80% accuracy and a macro-average ROC-AUC of 99.77%. Oracle, statistical, computational complexity, feature space, and Occlusion Sensitivity analyses were further conducted to assess model complementarity, computational efficiency, representation quality, and interpretability. The results demonstrate that feature-level fusion provides the primary performance improvement while OOF-based decision stacking offers complementary refinement within a leakage-aware evaluation protocol. Overall, the proposed framework emphasizes methodological rigor through leakage-aware meta-learning, independent testing, and comprehensive model analysis, rather than relying on classification accuracy alone. Full article

Underwater images are degraded by wavelength-dependent absorption and scattering, while enhancement on resource-constrained platforms requires a small model and low inference latency. We propose PAUNet, a lightweight encoder–decoder network combining partial axial depthwise convolution (PADW), multiscale axial spatial adaptive attention (MASA), and attention-guided skip connections. PADW filters half of the feature channels and preserves the remainder through an identity branch. MASA combines three axial scales using spatial branch attention. A joint objective constrains reconstruction, structure, perceptual features, and color. PAUNet achieves PSNR values of 21.508, 28.696, and 28.632 dB on UIEB, EUVP, and LSUI, respectively, with SSIM values of 0.751, 0.843, and 0.895. Among the compared methods, PAUNet has the highest PSNR on EUVP and LSUI, ranks second in UIEB PSNR, and achieves the highest LSUI SSIM. PAUNet contains 64,495 parameters, requires 0.493 GFLOPs, and runs at 288.95 frames per second on an NVIDIA RTX 3090. On an NVIDIA Jetson Xavier NX, TensorRT FP16 inference reaches 74.55 frames per second at 256 × 256 resolution and batch size one. The comparisons and ablations evaluate this compact architecture for underwater image enhancement. Full article

Aiming to overcome the challenges associated with the unmanned aerial vehicle (UAV)-based inspection of transmission line insulators, including the tiny scale of defect regions, irregular morphologies, and low contrast with complex backgrounds, we propose a lightweight insulator defect detection model based on real-time detection transformer (RT-DETR), termed FusionLite-DETR. To improve detection accuracy while maintaining computational efficiency, the proposed model introduces systematic enhancements to RT-DETR in four aspects: the backbone network, feature reconstruction operator, feature fusion module, and loss function. Firstly, the backbone network is reconstructed by integrating deformable large kernel attention (DLKA) into VanillaNet, enabling the model to couple global perception with precise local focusing, thereby significantly enhancing feature extraction for multi-scale defects. Secondly, the defect-aware content-aware reassembly of features (DA-CARAFE), is proposed to replace the conventional upsampling method. It dynamically generates reassembly convolution kernels based on underlying spatial features, achieving semantically aligned feature fusion under lightweight constraints. Meanwhile, a wavelet transform-based C3 (WTC3), is designed. By performing feature representation and interaction across different frequency sub-bands and incorporating ghost-shuffle convolution (GSConv), the module effectively reduces parameter redundancy while preserving critical high-frequency texture information. Furthermore, by introducing the inner minimum point distance intersection over union (Inner-MPDIoU) loss function, the boundary discrepancies are accurately captured by combining auxiliary-box overlap measurement with vertex-based geometric distance constraints, thus effectively mitigating the gradient vanishing problem for distant small targets during regression. FusionLite-DETR achieved mean average precision at an intersection-over-union threshold of 0.50 (mAP50) values of 95.5% and 93.1% on the private multi-scenario insulator defect dataset (PMID) and insulator fault detection dataset IFDD (datasets), respectively, with 19.1 M parameters and a computational cost of 54.8 giga floating-point operations (GFLOPs). Compared with RT-DETR with a resnet-50 backbone (RT-DETR-R50), the proposed model reduced the parameter count and GFLOPs by 54.8% and 57.4%, respectively, demonstrating a favorable balance between detection performance and model complexity. Full article

Hot-rolled strip steel surface defect detection is a vital industrial task that demands both high precision and real-time inference. Nevertheless, existing deep-learning-based detectors struggle to balance accuracy and speed, limiting their practical deployment on resource-constrained edge devices within harsh manufacturing scenarios. To address this gap, this paper develops YOLO-RMF, a real-time one-stage detector for hot-rolled strip steel inspection. Four core components are presented: a reparameterized 2 × 2 convolution (RPRC2) for enhanced feature extraction with lower inference computational cost than standard 3 × 3 convolutions, a two-dimensional axial attention (TDAA) module to capture long-range spatial dependencies, an adaptive-weighted multi-scale feature fusion pyramid network (MSF-FPN), and a hybrid adaptive pruning strategy combining BN scaling factors and feature-map rank analysis for model compression. Evaluated on NEU-DET and MVIT datasets, YOLO-RMF obtains 76.1% mAP50 and 40.6% mAP50-95 at 212 FPS (batch size 1, RTX 4090 GPU), achieving a favorable balance between detection sensitivity and precise localization for edge-computing-based industrial inspection. The pruning scheme reduces parameters by 36.3% and GFLOPs by 49.0% with minor accuracy loss. Deployed on Jetson AGX Orin, the pruned model reaches 169 FPS under FP16 and 233 FPS under INT8 quantization. Experimental outcomes show that YOLO-RMF achieves a favorable accuracy-speed trade-off and is suitable for practical industrial edge-based defect inspection scenarios. Full article

Background/Objectives: Objective assessment of patients’ physical health status before surgical interventions is crucial in predicting surgical risks and planning patient management. The American Society of Anesthesiologists Physical Status Classification (ASA-PS) is one of the most widely used scales in this context. It plays a vital role in clinical decision-making processes. However, ASA scoring can vary among clinicians due to its reliance on subjective assessments, making it difficult to standardize. Methods: In this study, a deep learning-based model was developed to classify ASA from patient data automatically. Three different data types were used in this study: lung images, Electrocardiogram (ECG) data, and clinical data. For the proposed ASA prediction model, features were extracted from each data type. Features were extracted from ECG and Computed Tomography (CT) images using five different Convolutional Neural Network (CNN) architectures, while statistical features were derived from the clinical data. The features obtained from the CNN architecture with the highest classification performance were combined with the statistical features. These combined features were then processed using seven different classifiers. Results: The developed model was evaluated with performance metrics such as accuracy, precision, F1 score, and Receiver Operating Characteristic (ROC) curve. As a result, a high accuracy value of 97.30% was obtained. Conclusions: This study aims to make preoperative risk management for ASA classification more objective and consistent, and the results obtained support this aim. Full article

The collection of solar radio burst spectrograms through the spatial distribution of survey sites reveals heterogeneity and differences in frequency sampling, temporal resolution, signal intensity, and local observing conditions. Because of this heterogeneity, it is difficult to incorporate observations of diverse stations directly into a single dataset, which restricts the generalizability of machine learning models that were trained using records from isolated stations. Thus, it is well established that for reliable automated analysis of solar radio bursts, a standard representation of spectrograms is necessary. This paper proposes a framework for harmonizing heterogeneous multi-site spectrograms of solar radio bursts, which transforms them into a unified representation capable of accommodating station-specific observational configurations suitable for machine learning applications. The proposed methodology combines adaptive frequency interval selection, frequency sampling consistency checking, logarithmic intensity transformation with baseline correction, time–frequency harmonization, and explicit handling of missing frequency channels using binary validity masks. To overcome the heterogeneous nature of the observations taken from different instruments and stations, nearest-neighbor interpolation in both time and frequency domains is applied. The unified frequency grid is constructed considering the statistics of the frequency channels used in spectrometers that form the e-CALLISTO network, a space weather instrument array. Two one-dimensional convolutional autoencoders with two latent space dimensions (32 and 64) are proposed and trained on more than 17000 spectrograms to evaluate the quality of latent representations of the harmonized data. The best performance was achieved by the autoencoder with simpler architecture and with 64-dimensional latent space, reaching root-mean-square error of $0.88$$\mathrm{dB}$ and a mean absolute error of $0.61$$\mathrm{dB}$. The experimental results show that models with 64-dimensional latent space outperform their 32-dimensional counterparts, while increased architectural complexity provided no substantial improvement. The proposed methodology provides a standardized basis for constructing harmonized datasets suitable for a wide range of subsequent machine learning applications in solar radio astronomy, including automated burst detection, classification, representation learning, anomaly detection, and predictive modeling. Full article

Human gait is a complex motor process regulated by the nervous system, where even subtle alterations in stride symmetry, stance duration or joint coordination directly reflect underlying motor and neurological impairments. Consequently, gait analysis provides a highly sensitive, objective, and non-invasive approach for clinical diagnosis and continuous monitoring. With the increasing global prevalence of neurological and motor impairments, there is a rising need for scalable objective methods capable of supporting diagnosis, monitoring progression and guiding rehabilitation. In this work, we present a Spatial–Temporal Attention-Guided Network (STAG-Net) for pathological gait classification using the GAIT-IT dataset, including five gait types associated with different problems: scissor, spastic, steppage, propulsive, and normal. The dataset provides a challenging benchmark characterised by class imbalance, high inter-class similarity, and significant intra-subject variability. To address these challenges, we develop a hybrid architecture combining Convolutional Neural Networks for spatial feature extraction with Recurrent Neural Networks to model temporal gait dynamics. We propose a Bidirectional Long Short-Term Memory module that captures gait dynamics, enabling the model to better represent cyclical and symmetrical gait characteristics. Furthermore, we integrate a temporal attention mechanism to dynamically weight salient gait phases, supporting the network’s ability to discriminate between visually similar gait abnormalities. The resulting STAG-Net framework achieves $96.82\%$ classification accuracy, demonstrating robustness in capturing gait differences across simulated pathological gait categories and highlighting its potential as a foundation for future extension toward real-world gait monitoring applications, pending validation on clinical patient populations. Full article

Facial expression recognition (FER) requires models that capture both fine local appearance and global facial structure. Convolutional neural networks provide strong local inductive bias, whereas vision transformers model long-range dependencies but can be data-intensive. This paper proposes EfficientHTNet, a compact dual-branch architecture that combines a partly fine-tuned EfficientNet-B0 local branch with a hierarchical Transformer global branch. A bidirectional cross-attention fusion module preserves and exchanges the 49 and 196 spatial tokens from the two branches, and an adaptive three-way gate combines the local, global, and interaction representations. The architecture is evaluated on FER2013 as the primary method-validation benchmark, on the synthetic SynPAIN corpus as a secondary breadth check, and on YT-DemTalk under subject-independent five-fold cross-validation as a proof-of-concept dementia-label classification task. The YT-DemTalk labels are derived from public videos and are not independently verified clinical diagnoses; therefore, those results must not be interpreted as Alzheimer’s disease diagnoses or clinical screenings. Across the evaluated tasks, the findings support a favorable accuracy, efficiency trade-off, while external validation, repeated runs, temporal modeling, and robustness analysis remain necessary. Full article

Retinal vessel segmentation is essential for diagnosing eye diseases. However, complex multiscale vascular structures and low contrast make accurate segmentation difficult. Current methods often overlook structural priors and seldom explore spatial–frequency synergies, resulting in poor segmentation of tiny vessels. To address these issues, we propose a Prior-Guided Spatial–Frequency Interactive Network (PGSFI-Net) that operates in both spatial and frequency domains. The Sobel-Attention Prior module extracts vascular edge priors using a single-scale Sobel operator and spatial attention, followed by multi-scale prior injection to strengthen faint-vessel representations. The Four-Branch Selective Attention Aggregation module captures multi-directional vascular features through strip convolutions and multi-dimensional attention. The Multi-Branch Spatial–Frequency Fusion module achieves bidirectional feature complementation via cross-gating interaction. The Local–Global Fusion module recovers subtle vascular structures by combining differential CNN branches with channel self-attention. On DRIVE, CHASE_DB1, and STARE, PGSFI-Net achieves ACC/F1/AUC values of 0.9646/0.8136/0.9838, 0.9727/0.8290/0.9838, and 0.9746/0.8393/0.9884, respectively. These results demonstrate competitive performance against state-of-the-art methods across the evaluated datasets and challenging low-contrast cases. Full article

Non-contact detection of weak, transient acoustic signals is essential for photoacoustic and laser-ultrasound (LUS) bioimaging, including the assessment of corneal deformation and thickness variation. Passive 3 × 3 fiber-coupler interferometers provide phase-diverse sensing without external modulation; however, conventional ellipse-based calibration becomes unreliable when the measured phase excursion covers only a small fraction of the characteristic ellipse, as in nanometer-scale corneal responses. This study presents a physics-guided temporal convolutional network (PG-TCN) that directly maps two nonideal interferometric outputs to a continuous dynamic phase signal without ellipse fitting or separate phase unwrapping during inference. The model combines dilated one-dimensional convolutions with interferometric-consistency and temporal-regularization losses. Tests using PZT-modulated calibration signals acquired under several static chamber pressures of 10, 15, 20, and 25 kPa showed lower mean absolute error, higher normalized cross-correlation, and lower displacement RMSE than a CNN-LSTM baseline. In corneal-phantom LUS measurements, the first detected echo shifted from 4.632 to 4.488 µs between 10 and 20 kPa, corresponding to a relative thickness reduction of approximately 7.8 µm/kPa. The PG-TCN required 210 ± 18 µs per sequence while using approximately half the trainable parameters of the CNN-LSTM baseline. These results demonstrate a fast physics-guided optical detection framework for non-contact sensing of weak acoustic responses, supporting the use of 3 × 3 optical interferometry as a biosensing approach for LUS and photoacoustic applications. Full article

Accurate ultra-short-term photovoltaic (PV) power forecasting is essential for the reliable operation and energy management of PV systems. However, rapid irradiance variations and complex meteorological conditions result in highly intermittent, nonlinear, and nonstationary PV power generation. Existing forecasting methods generally focus on temporal dependencies but provide limited representation of the dynamic relationships among meteorological variables and PV power. Moreover, conventional point-wise loss functions do not explicitly preserve the temporal shape of multi-step forecasts. To address these limitations, this study proposes a forecasting framework integrating a dynamic graph convolutional network (DGCN), long short-term memory (LSTM), and the transformation-invariant loss function with distance equilibrium (TILDE-Q). Meteorological variables and historical PV power are represented as graph nodes, and DGCN dynamically learns their sample-specific dependencies. LSTM captures temporal evolution patterns, while TILDE-Q constrains amplitude, phase, and scaling variations to improve trajectory consistency. The proposed method was evaluated on two independent operational PV datasets under 1-, 4-, 8-, and 16-step forecasting horizons, corresponding to 15 min, 1 h, 2 h, and 4 h ahead, respectively. DGCN-LSTM achieved the lowest RMSE and highest $R^2$ across all forecasting horizons on both datasets. Compared with the MAE training objective, TILDE-Q reduced RMSE and MAE by up to 15.35% and 25.90%, respectively, and consistently outperformed DILATE. These results demonstrate that combining dynamic variable-dependency modeling with shape-aware optimization can effectively improve the accuracy and trajectory consistency of multi-step PV power forecasts. Full article

In an era where IT security has paramount importance, the identification of devices holds a significance comparable to that of individuals. With the pervasive adoption of wireless technology, understanding the informational content within wireless data, particularly when intercepted by third-party devices, becomes crucial. This study aims to measure the effectiveness of a system that can capture Bluetooth signals from a device and, using these signals, create an image that can be used to identify the device with a neural network approach. Bluetooth packets were captured in an isolated environment, filtered, and transformed into images via post-processing Python code. A Convolutional Neural Network (CNN) based on a modified LeNet architecture was trained using 20,000 data samples per device. Testing with 36 different hyperparameter combinations showed accuracy above 95% for known device classes and approximately 40% confidence when presented with images from unknown devices. The results confirm that devices can be identified with acceptable accuracy based on Bluetooth signal fingerprints. Full article

Next point-of-interest (POI) recommendation seeks to predict the POI that a user will visit next from historical check-in records. Accurate recommendations require modeling POI relations and user trajectory patterns. Existing methods remain limited in representing continuous spatial influence decay, reliability-filtered temporal associations, semantic hierarchies, and multi-membership attributes. Furthermore, user trajectories may contain relatively persistent behavioral regularities and short-term variations within a check-in sequence, while standard time-domain models do not explicitly decompose trajectory representations by variation scale. Accordingly, this study introduces the Dynamic Hierarchical Attention Fusion Spatio-Temporal Network (DHAFST). Our Adaptive Dual-Graph Convolutional Network (ADGCN) employs Gaussian kernels with dataset-specific bandwidths to model continuous spatial influence and uses activity- and co-occurrence-based reliability filtering to construct temporal associations. The Bidirectional Hierarchical Graph Attention Network (BiHGAN) estimates soft memberships through a nested Gaussian mixture hierarchy and constructs cross-level affiliation relations for bidirectional hierarchical propagation. The Attention-Frequency Fusion Transformer (AFFT) combines time-domain self-attention with FFT-based low-frequency reconstruction and high-frequency residual modeling through an adaptive gate. Experiments on three real-world check-in datasets show that DHAFST achieves the highest Acc@1 among the evaluated methods on all three datasets. Full article