Search Results (1,035)

Real-time automatic detection of fire and smoke is a critical component of modern safety systems in surveillance, industrial, and environmental-protection applications, where conventional sensor-based systems show fundamental limitations. In this paper, a comparative analysis of four deep-learning architectures—DETR, Faster R-CNN (ResNet-50-FPN), Faster R-CNN (MobileNetV2), and RetinaNet—was conducted on a heterogeneous corpus of 67,765 annotated images originating from four different datasets. All models shared the same data-preparation, augmentation, and evaluation pipeline, while each was trained with the default configuration of its reference implementation; the comparison therefore reflects each architecture as it is typically deployed rather than a comparison under a single unified training budget. Faster R-CNN with the ResNet-50-FPN backbone achieved the highest accuracy (mAP@0.50 = 78.7% on the Indoor set; 73.8% on the SmokeAndFire set) and the highest mean mAP@0.50 across all datasets (48.7%), obtained with an inference speed of 76.9 FPS and a latency of 13.5 ms on NVIDIA RTX 4090 hardware, which makes it a promising candidate for real-time fire-detection systems on comparable hardware. The main contribution of this work is a unified evaluation of four representative object-detection architectures across four heterogeneous fire-and-smoke datasets. The study further quantifies the performance asymmetry between fire and smoke detection through a normalized morphological asymmetry index, reflecting the consistently lower detection accuracy achieved for smoke owing to its diffuse and semi-transparent appearance, and provides practical guidelines for selecting an architecture according to real-time deployment requirements. Because each configuration was trained once with a fixed random seed and all speed measurements were obtained on a single desktop GPU, the reported differences are interpreted descriptively; their statistical validation, cross-dataset evaluation, and measurement on embedded hardware are identified as future work. Full article

Maritime object detection is essential for port-based surveillance, ship tracking, and maritime security. However, practical implementation encounters two primary challenges: significant class imbalance (IMO identification numbers represent only 1.2% of labelled objects) and spatial inconsistency (predicted IMO numbers often appear beyond ship boundaries). Standard detectors and conventional class-balancing strategies fail to adequately address these difficulties, resulting in a persistent mismatch between research validation and practical performance. We present a hybrid autoencoder–YOLO framework with variable spatial regularisation. To the best of our understanding, this is the first methodology to simultaneously address class imbalance and spatial inconsistency in maritime IMO detection via reconstruction-guided learning and differentiable spatial regularisation. The model uses a YOLOv8 encoder shared by both a detection head (designed for ships and IMO numbers), and an auxiliary reconstruction decoder (a SkipDecoder with U-Net-style skip connections). A unique spatial limitation loss provides the physical restriction that each IMO number must be within a ship’s bounding box, using only anticipated boxes. Training occurs in two phases: autoencoder pretraining on the marine dataset, followed by phased joint optimisation with a curriculum schedule for the spatial weighting. In a dataset of 297 annotated images (utilising five-fold cross-validation with a 47-image preserved test set), our comprehensive model achieved 50.1% IMO AP50-95, 97.8% IMO precision, and 94.6% IMO F1-score on the test set, beating the baseline YOLOv8s by +7.8 percentage points, +6.5 percentage points, and +5.2 percentage points, respectively. Ablation studies indicate that reconstruction instruction improves IMO AP50-95 by +4.5 percentage points, while spatial regularisation adds +3.3 percentage points. Although ship detection results in a small compromise (ship AP50-95 decreases from 77.0% to 55.6%), this appears to be practically acceptable given the essential role of IMO numbers as unique ship IDs. Significantly, inference speed improves by 20% (8.0 ms per image on an NVIDIA A100 GPU) relative to YOLOv8s (10.0 ms). Comparisons with leading detectors (RetinaNet, Faster R-CNN, DETR, EfficientDet), all initialised with standard COCO-pretrained backbones and fine-tuned on our dataset, reveal that none achieve an IMO AP50-95 exceeding 42.3%, highlighting the task’s challenge and the accuracy of our design. The proposed framework shows potential to address the gap in practical implementation through reconstruction-based feature learning accompanied by a specified geometric baseline. It is precise, accurate, and fast, aligns with specific physics, and shows promise for real-time maritime surveillance applications, though we acknowledge the need for additional thorough verification across many operational environments. Full article

Continuous maritime surveillance increasingly relies on optical remote sensing, yet small ships are difficult to detect in low-resolution imagery because limited spatial resolution, complex sea backgrounds, and sensor noise suppress the fine-grained cues on which detectors depend, while routine acquisition of high-resolution imagery is prohibitively costly. To address this problem, a hybrid framework (RGT-YOLOv5Det) is proposed that couples transformer-based super-resolution (SR) with a lightweight object detector so that fine spatial detail is restored before detection. Methodologically, a paired benchmark (Ship-HRRSI/Ship-LRRSI) was constructed from the public TGRS-HRRSD dataset by standardising images to 800 × 800 pixels and generating 200 × 200 pixel counterparts via 4× bicubic down-sampling; three SR models (RGT, HAT, and Real-ESRGAN) and four detectors (YOLOv5s, YOLOv8s, YOLOv10s, and Faster R-CNN-MobileNetV3-Large-FPN) were fine-tuned and compared under identical training settings, and the best-performing components were integrated into the proposed two-stage pipeline. In the results, RGT delivered the best reconstruction quality (PSNR 22.038 dB, SSIM 0.3502) with the fewest parameters (13.37 M), YOLOv5s proved the most resolution-robust detector, and the integrated RGT-YOLOv5Det achieved mAP@0.5 of 0.947 and mAP@0.5:0.95 of 0.768 on low-resolution imagery, exceeding the best standalone detector score on each metric by 0.033 and 0.089, respectively. It is concluded that restoring structural detail prior to detection offers an accurate and acquisition cost-efficient alternative to high-resolution imaging, providing a practical route to reliable small ship detection in degraded optical remote sensing imagery. Full article

Generative artificial intelligence (GenAI) is diffusing faster than any recent technology, yet its net effect on energy use remains unsettled: the systems that enable smarter energy management are themselves energy-intensive. Using 22,143 firm-year observations on 3177 Chinese A-share listed firms over 2014–2023, this study measures corporate energy efficiency as a total-factor slacks-based measure score with undesirable output and dates GenAI disclosure from the fiscal year of a firm’s first annual-report mention, giving a staggered difference-in-differences design. Disclosure is followed by an increase in the efficiency score of between 6% and 9% of the sample mean, which is stable under parallel-trend, placebo and heterogeneity-robust checks and three alternative frontiers. Disclosing firms subsequently record more green innovation and report deeper digital transformation and more intensive use of data as a production factor. The post-disclosure association is larger where environmental regulation is tighter and digital infrastructure denser, and it is statistically significant among heavy-polluting and state-owned firms. In exploratory estimates, a firm’s book-measured artificial-intelligence capital intensity is separately associated with lower energy efficiency. Full article

In automotive body-in-white (BIW) inspection, small spot-weld defects can be obscured by metal-surface reflections and by the similar morphology of normal and defective welds. This study proposes FasterNet-Context-Guided Block-C2f-CNCM YOLO (FCC-YOLO), an improved YOLOv8n detector that applies stage-wise feature compensation. A FasterNet Block enhances local defect details in the backbone, a Context-Guided Block improves defect–background discrimination in the neck, and C2f-CNCM combines spatial, column, and self-calibration branches during multiscale fusion. On a dataset of 820 images (656 training, 82 validation, and 82 test images), FCC-YOLO achieved mAP50 of 96.0% and mAP50-95 of 74.7%, improvements of 3.0 and 3.2 percentage points over YOLOv8n. On an NVIDIA Jetson AGX Orin with TensorRT FP16, it processed images at 56.26 FPS with 17.77 ms mean end-to-end latency. These results demonstrate a favorable accuracy–efficiency trade-off in the evaluated dataset and deployment environment. Generalization across production lines, materials, and illumination conditions remains to be validated. Full article

Timely and accurate termite detection is essential for effective termite control. To address the challenges posed by the small size of termite individuals and the susceptibility of target features to background texture interference under complex backgrounds, this study proposes ESD-YOLO, a fine-grained feature-enhanced object detection model. Using YOLO11n as the baseline, ESD-YOLO redesigns the feature extraction, deep feature aggregation, and multi-scale feature fusion stages to improve the representation of small-scale termite targets under complex backgrounds. Specifically, the Efficient Multi-scale Attention (EMA) mechanism is incorporated into the C3k2 module to enhance feature discriminability between termite individuals and the background. A Spatial Pyramid Pooling-Fast with Dual Global Pooling (SPPF-DGP) module is employed to supplement deep features with global contextual information and salient response information. In addition, the DySample dynamic upsampling module is introduced to improve spatial alignment during multi-scale feature fusion and enhance boundary representation for small targets. Experimental results show that ESD-YOLO achieves Precision, Recall, mAP@0.5, and mAP@0.5:0.95 values of 95.39%, 96.33%, 97.85%, and 65.92%, respectively, with 2.67 M parameters and 6.68 G FLOPs. Compared with Faster R-CNN, RetinaNet, RT-DETR, and several YOLO-series models, ESD-YOLO demonstrates strong small-target detection and localization performance under the controlled complex-background conditions established in this study, providing a methodological reference for automated termite detection in practical settings. Full article

Falls are a major cause of severe injury and mortality among older adults, requiring rapid state recognition and alerts in healthcare and caregiving environments. However, many existing fall-recognition approaches primarily focus on post-fall detection or assume access to cloud/GPU computing, leaving limited evidence for short-term pre-fall recognition on low-end CPU-based edge devices. This study proposes a lightweight CNN–TCN–Attention framework for recognizing fall and pre-fall states on resource-constrained edge devices. Using MediaPipe, three-dimensional coordinates and visibility scores of 33 human body landmarks are extracted from input videos. Raw RGB frames are not directly used by the classifier, thereby reducing the processing of visually identifiable information. A total of 2300 fall-related videos from the AI-Hub dataset were reorganized into three classes: normal, pre-fall, and fall. The pre-fall onset was defined as the point at which the downward vertical velocity of the nose landmark exceeded a dataset-specific threshold. The model uses a CNN to extract frame-level skeletal patterns, a temporal convolutional network (TCN) to learn temporal changes in posture, and an attention module to aggregate temporal features. It was deployed on a Raspberry Pi Zero 2 W using ONNX Runtime. The proposed model achieved 94.71% accuracy and a macro F1-score of 93.63%, with an average state-decision latency of 325.88 ms/decision. It improved accuracy by 2.83 percentage points over the lightweight CNN–LSTM–Attention baseline while maintaining comparable latency and achieved approximately 1.75× faster processing than the MobileNetV3–TCN–Attention model. Full article

Cattle producers must balance animal welfare with productivity, but regular observation of animals in extensively managed operations is often impractical. Artificial intelligence (AI) integrated with computer vision offers an automated alternative, but few studies compare object detection architectures within the same dataset or focus on pastured cattle. Groups (n = 2–7) of heterogeneous beef cattle were recorded on pasture with nine solar trail cameras. Footage (~132 h) was curated in VideoLAN; annotated in Computer Vision Annotation Tool (CVAT) with bounding boxes and behavioral classes; and split 62/21/17% into training (24,508 frames), validation (8231 frames), and testing (6625 frames) sets. Four architectures were trained: Faster R-CNN (ResNet-50 FPN), Single Shot MultiBox Detector (SSD300, VGG-16), RetinaNet (ResNet-50 FPN with focal loss), and YOLOv8 nano (Ultralytics). With validation at 0.50 confidence and 0.50 IoU, Faster R-CNN achieved the highest overall F1 (0.79) and best per-class balance; RetinaNet was intermediate (peak F1 = 0.72); SSD300 saturated at F1 = 0.40; and YOLOv8 nano achieved some minority class recall at lower confidence. Each model detected the classes “grazing” and “hay feeding” accurately but confused cattle with the visually similar “normal” class. Datasets, checkpoints, and analysis scripts are provided to support further refinement of AI-enabled monitoring of extensive cattle systems. Full article

The growing frequency of unauthorized UAV activities has increased the demand for real-time perception and rapid response on resource-constrained edge devices. This study proposes an edge-deployable UAV detection and net-capture system based on Net-Capture Detection YOLO (NCDet-YOLO). Developed from YOLOv8n, NCDet-YOLO incorporates C2f_Faster, SPD_Conv, EMA, and a lightweight three-scale detection head, with CrossKD used to compensate for accuracy loss caused by structural compression. The dataset contains 6615 images and was divided into 5292 training and 1323 validation images. The self-collected data include DJI Phantom 4 and DJI Inspire 2 UAVs observed at approximately 4–30 m under different daytime backgrounds. NCDet-YOLO achieves an mAP50–95 of 0.6504 with 1.55 M parameters and 4.1 GFLOPs. On a Jetson Orin NX Super under the 15 W power mode, it achieves 31.53 FPS, representing a 31.67% increase over YOLOv8n. The detector is further integrated with target alignment, distance determination, trigger control, and net-capture execution. In 10 real-platform trials, 8 captures were successful, corresponding to an 80.0% success rate, with one false-trigger event and an end-to-end latency from target detection to net-capture firing of approximately 400 ms. Full article

Thermite mixtures are attractive for downhole well plug-and-abandonment (P&A) sealing, where a consistent, controllable burn matters more than peak energy. Because propagation is governed by particle size and packing, predictive blend design needs a model that resolves microstructure. I present a zero-tuning multiscale framework that links a discrete element method (DEM) packing directly to combustion-front behaviour. The DEM contact graph sets a contact-resistance effective conductivity (a thermal-network solve on the contacts, cross-checked against a packed-bed model); the particle sizes set a diffusion-limited, product-layer (shrinking-core) reaction rate with the measured activation energy of the Al-Fe₂ O₃ reaction (Eₐ = 145 kJ mol⁻¹); and these feed an analytical condensed-phase travelling-wave speed that is mesh-free by construction and confirmed against a converged numerical eigenvalue solve. Here “zero-tuning” means no coefficient is fitted to the blend dataset: every transport and kinetic parameter is DEM-derived or taken from the literature, and a single diffusion pre-factor is anchored to an independent fine-powder benchmark. Applied to a generic Fe₂ O₃/Al+ sand system across eight coarse (∼256–462 µm) +40/+70 blends, the framework predicts front speeds of ∼2–4 mm/s—about an order of magnitude below fine powders (27–47 mm/s)—i.e., finer-is-faster, as expected for diffusion-controlled aluminothermic reactions; the residual size dependence is the net of competing diffusion-kinetic and radiative effects rather than a clean monotonic lever. A DEM-derived Kozeny–Carman permeability shows gas convection contributes ≲15% of the front enthalpy, justifying the conduction–radiation formulation. The DEM-microstructure-to-rate mapping is validated externally on Ni–Al self-propagating high-temperature synthesis (SHS), whose measured particle-size ordering it reproduces. The contribution is the framework itself—a predictive, microstructure-resolved route requiring no blend-specific fitting. The computational implementation used a supervised AI-agent workflow (implementation, execution, verification); all scientific content was conceived and verified by the author. Full article

Transient contact responses in confined guide channels contain sharp events and parameter-dependent phase shifts, which make fixed-output full-history surrogates difficult to train. This study develops a time-conditioned residual surrogate (TC-ResNet) to propagate uncertainty in the bounded center-of-mass eccentricity components into the transient acceleration responses of a generic pressure-driven, high-overload axisymmetric body. The nonlinear reference model includes prescribed base pressure, wall contact and impact, velocity-dependent friction, gravity, pitch and yaw, and eccentricity. TC-ResNet predicts one response value for each parameter–time query by combining normalized physical parameters with Fourier-embedded time. The axial acceleration is modeled directly, whereas low-frequency radial trends and sliding root-mean-square (RMS) curves represent dominant lateral motion and local vibration intensity. On a common 108-case test set, TC-ResNet achieved the highest coefficient of determination (R²) for a_x^low (0.850), a_x^rms (0.839), and a_z^low (0.801), as well as the highest macro-mean R² (0.848). The fixed-output multilayer perceptron (MLP) remained best for a_y (0.995) and a_z^rms (0.775), demonstrating that the proposed model is not uniformly superior. Interval analysis shows that the prescribed pressure load limits axial sensitivity, whereas radial offsets alter eccentric pressure moments and wall contact; the axial offset primarily changes contact and friction moment arms, and inclination mainly affects the later radial response through gravity decomposition and accumulated contact differences. Accuracy approaches a plateau between 378 and 504 training cases. On the same central processing unit (CPU), TC-ResNet requires 0.081 s per curve (approximately 170× faster than the reference solver), and graphics processing unit (GPU) inference requires 0.019 s per curve. The reported envelopes support qualitative sensitivity analysis, but experimental calibration and validation remain necessary. Full article

Introduction: Pediatric clavicle fractures are among the most common fractures in this age group. Current research supports the efficacy of nonoperative treatment for these injuries in pediatric patients. Our study seeks to examine the current trends in the management of these fractures and determine whether they align with the existing evidence-based guidelines. Methods: A retrospective analysis of pediatric patients with clavicle fractures between 1 January 2014 and 31 December 2024 was analyzed. A retrospective query was performed using the TriNetX Research Network. Patients were dichotomized into two groups based on open reduction internal fixation and nonoperative or closed treatment codes over this period, performing subgroup analyses by sex, race, and age. Results: We evaluated 78,251 patients between 10 and 18 years with clavicle fractures from 2014 to 2024, of whom 3546 underwent open reduction and internal fixation (ORIF). ORIF utilization increased from 1.62% in 2014 to 5.89% in 2024, representing a 263.6% increase from the 2014 baseline. Males consistently underwent ORIF more often than females. Adolescents (15–18 years) had substantially higher and faster-rising ORIF rates than younger patients (10–14 years), increasing from 3.4% to 12.4% versus 0.8% to 2.5% between 2014 and 2024 (APC 7.5% vs. 0%, p = 0.001). White patients had higher ORIF rates than African American patients, although some disparities narrowed over time. Conclusions: Despite limited supporting evidence, current trends demonstrate a substantial increase in the surgical management of pediatric clavicle fractures, with operative treatment rising by 263.6% by 2024. This increase in ORIF procedures may expose patients to potentially avoidable risks, including infection, anesthesia-related complications, implant-related issues, and prolonged recovery, without clearly established benefits in many cases. Full article

Artificial intelligence (AI) data centers are creating large, power-dense loads, often faster than transmission lines, substations, transformers, and grid interconnections can be expanded. This review assesses whether solid oxide fuel cells (SOFCs) can provide dependable on-site power during these grid delivery constraints and remain competitive after grid capacity becomes available. We critically synthesized evidence on AI electricity demand, competing power supply options, SOFC efficiency and durability, commercial deployments, environmental impacts, thermal and electrical integration, and hybrid SOFC–battery–grid systems. We also performed a screening-level levelized cost of electricity sensitivity analysis covering natural gas prices, carbon costs, stack replacement, grid electricity prices, and the avoided cost of delayed grid access. The evidence indicates that commercial SOFC systems can achieve approximately 50–60% net electrical efficiency and scale modularly from 325 kW units to a planned deployment of up to 2.45 GW. A nominal 100 MW installation would require approximately 308 such modules and at least 3600 m² of direct equipment area, excluding auxiliary systems and safety setbacks. However, multi-year durability targets of about 40,000 h, fuel and carbon price exposure, slow transient response, lifecycle methane emissions, and limited opportunities to use high temperature exhaust heat remain important constraints. The economic analysis indicates that avoided grid delay costs can justify SOFCs as bridge assets, whereas long-term retention requires competitiveness without this temporary benefit. SOFCs are therefore most suitable for sites that prioritize rapid access to firm power, modularity, reliability, and low local air pollutant emissions, rather than as a universal alternative to grid expansion. Full article

Accurate identification of mango varieties holds substantial significance for the elevation of product added value and the facilitation of market differentiation through quality-based pricing. Near-infrared (NIR) spectral analysis offers a rapid, non-destructive solution for mango variety identification. To address the challenges in fine-grained classification of NIR spectra, namely, high spectral similarity and severe overlap of absorption peaks, which make it difficult to extract nonlinear features using chemometrics, as well as the excessive complexity of existing deep learning models, a lightweight multi-scale spatial global context network is proposed. One-dimensional NIR spectra are converted into two-dimensional images through the Gramian angular difference field. Multi-scale partial convolution, coordinate-aware global context, efficient multi-scale attention, and structural re-parameterization are integrated to capture local spectral features and long-range band correlations effectively. Evaluated on two mango spectral datasets with different distributions, the proposed model achieves variety identification accuracies of 99.46% and 97.83%, with only 19.08 M parameters. Computational complexity, throughput, and latency reach 120.29 M FLOPs, 2848.5 FPS, and 0.351 ms, respectively, realizing a balance between classification accuracy and computational speed. Ablation and robustness experiments demonstrate that the accuracy of the model is improved by 5.91% and 2.15% compared with one-dimensional convolutional neural network and FasterNet, respectively. Important wavelengths obtained by threshold screening of activation maps exhibit consistency with the majority of conclusions from analysis of variance and VIP methods, while the remainder represent newly identified important bands. Validation across different temperature and batch scenarios reveals strong generalization capability. Future refinement will be pursued through increased sample diversity. Overall, high-precision identification is attained by the model at comparatively low computational overhead, indicating potential for advancing the practical application of NIR spectroscopy in agricultural quality inspection. Full article

Identifying plant health conditions is an emerging precision-agriculture and food-security challenge, intensified by deploying deep-learning models on memory- and power-constrained edge devices. We present a TinyMLOps pipeline spanning model design, optimization, quantization, and deployment across diverse edge devices, evaluated under controlled laboratory conditions. Using a dataset derived from the PlantVillage benchmark, 39 fine-grained classes are aggregated into three superclasses: healthy leaf, unhealthy leaf, and no leaves. The resulting system therefore performs plant health-status classification and background filtering rather than diagnosing specific diseases. We train a MobileNet-based convolutional neural network jointly optimized for classification accuracy and computational efficiency, adopting state-of-the-art hyperparameter optimization (HPO) tools. Five models are selected, four from the Pareto Front and one as the biggest evaluated model during HPO, converted to LiteRT and ONNX, and evaluated at float32 and post-training int8 precision on a Raspberry Pi Zero 2 W and an STM32H743ZI microcontroller. At float32, LiteRT is 1.87–2.65× faster than ONNX Runtime on the Raspberry Pi across all five models. Relative to their float32 LiteRT counterparts, the int8 LiteRT models are 2.83–3.67× smaller on disk and 21.3–31.2% faster on the same board, at a cost in F1-score of between 0.0010 and 0.0068. On the microcontroller, only the two smallest models deploy at both precisions; for these, the fully quantized int8 variants are $4.3\times$ faster and 3.85× smaller in MCU flash footprint than the float32 counterparts. The mid-range model fits the 2 MB flash and 1 MB RAM budget only when quantized, while the two largest models exceed it in every configuration tested. Full article