Search Results (1,211)

Computer vision with artificial intelligence (AI) offers a promising tool for blueberry growers to accomplish orchard tasks such as harvest maturity assessment and yield estimation, which otherwise would be labor-intensive and prone to error. However, blueberry detection in natural environments remains challenging due to variable natural lighting, frequent occlusions by leaves and branches, and motion blur due to environmental factors and imaging devices. AI models such as deep learning-based object detectors promise to address these challenges, but they are data-driven, demanding a large-scale, diverse dataset that captures the complexities of real-world orchard conditions. Deployment of these models in practical scenarios often faces limited computing resources, highlighting the importance of achieving the right accuracy/speed/memory trade-off in model selection. This study presents a novel comparative benchmark analysis of advanced real-time object detectors, including YOLO (You Only Look Once) (v8–v12) and RT-DETR (Real-Time Detection Transformers) (v1–v2) families, consisting of 36 model variants, evaluated on a newly curated large dataset for blueberry detection. This dataset contained 661 canopy images collected with smartphones during the 2022–2023 seasons, consisting of 85,879 manually annotated instances (including 36,256 ripe and 49,623 unripe blueberries) that represent a broad range of lighting conditions, occlusions, and fruit maturity stages. Among the YOLO models, YOLOv12m achieved the best accuracy with a mAP@50 of 93.3%, while RT-DETRv2-X obtained a mAP@50 of 93.6%, the highest among all RT-DETR variants. The inference time varied with the model scale and complexity, and the mid-sized models appeared to offer a good balance between accuracy and speed. To further improve fruit detection performance, all models were fine-tuned using Unbiased Mean Teacher-based semi-supervised learning (SSL) with 1644 cross-source unlabeled canopy images acquired from ground-based machine vision platforms. SSL resulted in accuracy improvements of up to 2.0%, with RT-DETR-v2-X achieving the highest mAP@50 of 95.5%. These findings highlight the efficacy of SSL for leveraging cross-domain unlabeled data, although further research is needed to fully exploit its benefits. The curated dataset and developed software programs are publicly available to facilitate further research and practical deployment. Full article

Steam injection is an energy-intensive enhanced-oil-recovery method for high-viscosity reservoirs, and its performance is controlled by coupled heat delivery, steam condensation, temperature-dependent viscosity reduction, mobility change and reservoir filtration response. This study develops an experimentally validated physics-motivated residual-learning framework for forecasting oil production and selecting thermally rational steam-supply regimes. The model combines a physics-motivated semi-empirical baseline describing useful steam-related heat input, calibrated viscosity transformation, mobility growth, steam–oil ratio and a thermal-energy efficiency index with a residual-learning block fitted to measured regime-level records. The supervised forecasting task was performed at the regime level using 200 operating-regime records treated as the effective modelling units, with nested logs aggregated within regimes and within-group dependence examined through campaign-, reservoir-state- and well-availability-based checks. The 4800 steam-injection log entries and 4800 production-response log entries were treated as nested time-resolved measurements used only for regime-level aggregation, feature construction and quality-control checks; they were not counted as independent training samples. Blind-test validation produced R² values of 0.974 for oil rate, 0.988 for cumulative oil production, 0.731 for steam–oil ratio and 0.828 for the thermal-energy efficiency index; the corresponding MAPE values were 4.56%, 3.86%, 4.48% and 3.29%, respectively. The error structure shows higher uncertainty for composite indicators than for direct production responses, which is consistent with the measurement chain. Response-surface and Pareto analyses identify bounded steam-supply operating regions where production gain remains balanced against specific steam consumption and the thermal-energy efficiency index. Full article

Multi-view learning can effectively exploit the consistency and complementarity among multiple data sources and has become a major research direction in semi-supervised classification. However, the existing methods commonly suffer from several limitations, including the loss of view-specific information caused by premature feature fusion, interference from redundant inter-view noise, and the limited discriminative capability of consensus representation. These issues severely restrict classification performance under low-label settings. To address these limitations, this paper proposes Dual-branch Multi-view Learning with Dual-contrastive Information Bottleneck. The proposed framework constructs a decoupled dual-branch graph convolutional architecture to explicitly separate view-specific representations from cross-view consensus representation, thereby alleviating feature homogenization at the structural level. Furthermore, we design a dual-contrastive information bottleneck optimization mechanism, where the CLUB constraint minimizes redundant mutual information across views to suppress noise, while the InfoNCE constraint maximizes the mutual information between consensus representation and labels to enhance discriminative capability. Additionally, we employ an adaptive attention fusion module to dynamically integrate the dual-branch representations, further refining task-relevant features. The experiments conducted on nine public datasets demonstrate that the proposed method achieves favorable performance improvements over most of the selected comparison methods in semi-supervised classification tasks. Full article

Background/Objectives: Conjunctival hyperemia is a common clinical finding in clinical practice; however there are significant differences between graders. Vessel detection using deep-learning approaches could enable more objective measures. We aimed to evaluate vascular metrics derived from automated vessel detection and compare these metrics with manual severity gradings. Methods: Slit lamp images from 139 glaucoma patients were included. Images from 103 participants were used as the primary development dataset and the remaining as a validation subset. The images were independently graded by two graders for conjunctival hyperemia using the Efron Grading Scheme. Conjunctival vessels were detected using an automated vessel detection pipeline based on semi-supervised learning. Vessel density, fractal dimension and tortuosity were calculated and compared with the manual Efron grades. Results: Grading of conjunctival hyperemia between the two graders were consistent (Spearman’s rho: 0.79; ICC: 0.79 [95%CI: 0.72–0.84]) but showed significant differences with a higher proportion of differences in the moderate grades. Of the vascular metrics, vessel density showed significant associations with the individual Efron grading and against the mean Efron grading (0.78, p < 0.001). Fractal dimension was significantly associated with the mean Efron grading (0.55, p < 0.001). Agreements were similar in the subset (vessel density, 0.80, p < 0.001; fractal dimension 0.62, p < 0.001). Vessel tortuosity showed lower agreements (<0.23). Conclusions: Vessel density and fractal dimension showed significant associations with manual Efron gradings. These metrics could be potentially used to enable more objective and interpretable measures of conjunctival hyperemia severity. Full article

Graph Neural Networks (GNNs) have achieved strong performance on graph-structured data via neighborhood message passing. Recent studies on GNNs suggest that not all feature dimensions benefit equally from message passing, motivating preference-guided feature splitting rather than uniform aggregation. Empirically, the splitting criterion is affected by class-boundary nodes with label-inconsistent neighborhoods, which confound the estimation of which dimensions should be propagated. Moreover, conducting propagation on the original topology may amplify feature–topology mismatch, causing messages to be passed along incompatible edges. To address these issues, we propose a plug-and-play architecture called Core–Shell Adaptive augmentation with dual-branch Polarized Topology Refinement (CSA-PTR), through simultaneous consideration of clearer feature splitting and a more ideal topology for GNNs to better satisfy the selective propagation criterion. Specifically, CSA-PTR consists of three modules. Core–shell adaptive augmentation stabilizes node representations by a purity-aware clustering algorithm, which reduces the ambiguity in feature-preference estimation. Then, graph feature splitting allocates feature dimensions into a propagation branch and a feature-only branch based on learned preferences. Finally, Dual-branch Polarized Topology Refinement exploits these branches as complementary views to learn polarized weights, yielding a more desirable topology and improving information flow. Extensive experiments on diverse benchmarks show that CSA-PTR achieves competitive performance across the evaluated settings, while consistently improving several representative GNN backbones. Full article

Low-resource hate speech detection remains a challenging task for Indonesian social media due to limited labeled annotations, highly informal linguistic expressions, and substantial lexical variability. Under such conditions, purely supervised transformer models often suffer from unstable semantic generalization, while conventional pseudo-labeling methods are vulnerable to noisy unlabeled sample propagation. To address these limitations, this study proposes a hybrid feature-enhanced IndoBERT framework integrated with a controlled semi-supervised learning strategy. The proposed model combines contextual IndoBERT embeddings with abusive lexicon cues, handcrafted linguistic indicators, and TF-IDF–SVD statistical representations through a lightweight concatenation–projection feature fusion mechanism, while unlabeled data are incorporated via adaptive confidence thresholding and class-balanced pseudo-label selection to improve pseudo-label reliability. Extensive experiments were conducted under realistic low-resource supervision settings using only 5%, 10%, and 20% labeled data, and the proposed framework was systematically compared against representative baselines, including sparse lexical machine learning models, shallow neural architectures, multilingual transformers, IndoBERTweet, naive pseudo-labeling, and LLM-based prompting. The results show that model effectiveness is strongly supervision-dependent. Under the most extreme low-resource setting, compact statistical augmentation provides the most stable complementary signal, whereas under moderate low-resource supervision, the full hybrid representation combined with controlled semi-supervised learning yields the strongest and most consistent gains. The proposed Hybrid IndoBERT + controlled SSL framework outperforms all baselines at the 20% labeled setting, reaching an accuracy of 0.8654, Macro-F1 of 0.8633, and ROC-AUC of 0.9334. Additional analyses of pseudo-label reliability, calibration behavior, computational efficiency, and qualitative error patterns further show that the proposed framework improves low-resource robustness while maintaining comparable inference-time efficiency. These findings demonstrate that low-resource hate speech detection benefits most from the staged integration of contextual semantic modeling, interpretable linguistic cues, global lexical–statistical structure, and carefully regulated unlabeled data exploitation. Additional experiments using GPT-4o-mini and Llama-3.1-8B further demonstrate that the proposed framework remains competitive against general-purpose large language model prompting approaches under low-resource Indonesian hate speech detection scenarios. The proposed framework provides a practical and reproducible direction for hate speech detection in annotation-constrained social media environments. Full article

Mineralogical data acquired from drillcores provide important constraints for resource estimation, geometallurgical modelling, mineral exploration, and geological interpretation. While hyperspectral imaging is rapidly gaining traction for these applications, it lacks the ability to accurately quantify mineral abundances without extensive calibration data. Here, we build on previous work to demonstrate and benchmark workflows that combine scanning electron microscope (SEM) mineral maps with large-extent multimodal hyperspectral imaging data. The goal is to relate hyperspectral features and mineral abundances using supervised machine learning models, and then apply these models to infer mineralogy across entire drillcores. We adapt the learning process to the non-uniform (unbalanced) composition of most rocks, and achieve reasonable accuracy for most rock-forming minerals. However, we also find that prediction accuracy depends strongly on the representativity of training data—so models often fail to produce accurate maps of rare and accessory minerals. Robust, adaptive and ideally semi-automated sampling approaches might address this shortcoming by identifying locations which ensure optimal coverage of hyperspectral variance. We also emphasise that upscaling from SEM to drillcore scale inevitably involves extrapolation, meaning predictions should always be validated. However, once validated, upscaled mineralogy predictions could provide crucial quantitative data that bridge the scale gap between petrographic observations, metallurgical tests, and geometallurgical models. Full article

Currently, clustering algorithms are mainly used to classify fiber-reinforced composite cylinder damage. However, the number of clustering categories is heavily influenced by the evaluation criteria, and the real damage type categorization cannot be determined. Therefore, we propose a semi-supervised algorithm that obtains higher damage classification information with a small number of labels. Specifically, we first performed a phased fiber-reinforced composite cylinder pressurization experiment and collected damage signals through acoustic emission (AE) hits. We analyzed the damage types of the collected burst-type acoustic emission hits (each hit corresponds to a single waveform captured when the hit’s amplitude exceeds the preset threshold) and marked a small number of these hits. Then, we constructed a mean-teacher semi-supervised network structure based on transfer learning, achieving a classification accuracy of 85.92%. Compared to traditional supervised learning and clustering algorithms, the accuracy improved by nearly 30%. Full article

Power line component inspection is essential for maintaining the reliability of the electrical power infrastructure. Recently, some researchers have studied automatic power line inspection using drones and deep learning. However, fully supervised deep learning approaches require large amounts of labeled data that are difficult and expensive to obtain in real-world environments. To address these challenges, this paper proposes an edge-optimized semi-supervised deep learning framework for power line component inspection. The proposed approach combines a semi-supervised learning (SSL) strategy to leverage both limited labeled images and abundant unlabeled field data with hardware–software (HW-SW) co-optimization techniques for efficient deployment on resource-constrained edge devices. In the learning stage, the framework improves detection performance by leveraging unlabeled inspection data via pseudo-labeling and confidence-based sample selection, thereby reducing annotation effort while maintaining robust recognition performance. In the deployment stage, the quantization technique was applied to enable real-time operation on embedded platforms with limited computational resources and power budgets. In this paper, an improved version of the edge-AI deployment score, the generalized edge-AI deployment score (GEADS), is proposed. In SSL evaluation, debiased semi-supervised learning (DeSSL) achieves a higher observed mAP@0.5 and F1-score than the standard SSL method in the single-run simulations using dataset 1 and dataset 2. In hardware evaluation, the YOLOv7-Tiny (INT8) configuration implemented on a Raspberry Pi 5 achieves the highest GEADS of 0.657, confirming it offers the most balanced performance among the required parameters. From the simulation, it is also confirmed that the proposed GEADS provides a more interpretable and statistically stable metric than the existing metric to evaluate the edge deployment. Full article

Modern connected vehicles rely on the controller area network (CAN) to disseminate safety-critical in-vehicle information, including sensor-related and vehicle-state signals such as engine revolutions per minute (RPM) and gear state, among electronic control units (ECUs). Because CANs lack built-in authentication and encryption, malicious message injection and spoofing can compromise the integrity and availability of vehicular sensing and control functions. Existing deep-learning-based intrusion-detection systems (IDSs) show a clear trade-off: supervised methods perform well on known attacks but rely on costly labels, whereas unsupervised methods can identify unseen attacks but often suffer from high false-positive rates. To address these limitations, this paper proposes a semi-supervised generative adversarial network (SGAN) framework for CAN bus intrusion detection that combines image-based CAN representation with adversarial learning. Consecutive CAN messages are converted into $64\times 9$ grayscale images, and the proposed framework is trained in three phases. First, the discriminator establishes an initial decision boundary using a small labeled subset. It then refines this boundary through distribution-level likelihood objectives and generated samples. Finally, the generator is trained to produce realistic samples capable of deceiving the discriminator. The proposed method was evaluated on the Hacking and Countermeasure Research Lab (HCRL) car-hacking dataset using leave-one-class-out experiments to simulate unknown attacks and achieved an average accuracy of 99.73% and an average F1-score of 99.63% on unknown attacks. Moreover, with only 0.21 M parameters and 3.25 M floating-point operations (FLOPs), the model is well suited for resource-constrained in-vehicle platforms. These results indicate that the proposed framework can serve as a practical cybersecurity component for protecting CAN-carried data in vehicular sensing applications. Full article

Obtaining high-quality labeled data for supervised learning is costly, motivating the use of crowdsourcing, which distributes the annotation process across multiple workers with varying levels of expertise. A key challenge in crowdsourced data is annotation sparsity, as each worker labels only a limited subset of instances. This sparsity can amplify class imbalance, reduce supervision for minority classes, and bias standard cross-entropy-based models toward the majority classes. To address this problem, we propose a correlated chained Gaussian process framework trained on a focal-loss-based variational objective (CCGPFL). This probabilistic framework jointly models latent ground-truth and instance-dependent annotator reliability while accounting for correlations among annotators. In addition, the focal-weighted objective mitigates the imbalance induced by sparse annotations by assigning greater importance to harder examples during training. Experiments on synthetic, semi-synthetic, and fully real multi-annotator datasets show that CCGPFL achieves competitive and often superior performance relative to state-of-the-art learning-from-crowds baselines in terms of Overall Accuracy (OA) and Area Under the ROC Curve (AUC). Full article

Accurate brain tumor segmentation from magnetic resonance imaging (MRI) is essential for computer-assisted diagnosis, treatment planning, and disease monitoring. However, brain tumors usually exhibit irregular, heterogeneous, and multi-scale spatial patterns with complex and ambiguous boundaries. At the same time, the performance of deep segmentation models is often constrained by the limited availability of voxel-level annotations, which are expensive and time-consuming to obtain. To address these challenges, this paper proposes Semi-SwinUNeTR, a semi-supervised framework for 3D brain tumor segmentation with limited annotated data. The proposed method adopts SwinUNeTR as the segmentation backbone, enabling hierarchical volumetric representation learning through shifted-window self-attention while preserving the encoder–decoder structure required for dense prediction. On top of this backbone, we introduce a dual-consistency semi-supervised learning strategy, consisting of mean teacher-based model consistency and interpolation consistency-based data consistency. In addition, voxel-wise consistency weights are used to redistribute semi-supervised supervision toward structurally complex and boundary-irregular tumor regions without changing the SwinUNeTR backbone. Experiments on the BraTS 2019 benchmark demonstrate that the proposed framework achieves strong performance across different annotation ratios. The original Semi-SwinUNeTR achieves Dice scores of 84.93%, 86.25%, 87.05%, and 87.83% under the 10%, 20%, 40%, and 80% labeled-data settings, respectively. With the weighted consistency extension, the Dice scores are further improved to 85.64%, 87.94%, and 88.59% under the 10%, 20%, and 80% labeled-data settings, respectively, while the corresponding HD₉₅ values are reduced to 8.9826, 8.1854, and 7.4533. These results indicate that combining a SwinUNeTR backbone with complementary model consistency, data consistency, and voxel-wise consistency weighting is an effective strategy for semi-supervised volumetric medical image segmentation under limited annotation. Full article

All-in-one weather-degraded image restoration aims to restore clean images from diverse weather-degraded observations (such as rain, haze, and snow) using a unified model. However, this topic remains challenging due to its ill-posed nature and the scarcity of large-scale paired training data. This article develops a novel semi-supervised learning framework, termed Semi-Supervised Degradation-Aware Learning (S²DAL), to adjust the feature space to align with the unified parameter space for all-in-one adverse weather removal. Specifically, the proposed S²DAL consists of two backbone networks: a Degradation-guided Histogram Transformer (DHformer) for weather-degraded image restoration and a Degradation-guided Convolutional Neural Network (DCNN) for degradation generation. A key component, the Degradation-guided Histogram Transformer (DHT) block, is designed to effectively capture intrinsic image features while suppressing diverse degradation interference through channel shuffling modulation, dynamic-range histogram self-attention, and dual-scale gated feed forward. Furthermore, a Monte Carlo-based Expectation-Maximization (EM) algorithm is introduced to jointly optimize latent variables and network parameters under both labeled and unlabeled data. Extensive quantitative and qualitative results on synthetic and real-world datasets consistently demonstrate that the proposed S²DAL achieves superior restoration performance compared to multiple state-of-the-art fully supervised and semi-supervised approaches. Full article

This paper investigates machine learning (ML)-supplemented workflows integrated within a modular parametric modelling framework derived from a typological analysis of stadiums. The objective of the research is to address a gap between numerous isolated computational studies and the realities of early stadium design within the Royal Institute of British Architects (RIBA) Plan of Work (PoW) Stages 0–3. From a practical perspective, the proposed design framework aims to embed supervised learning, semi-supervised learning, and evolutionary optimisation into stadium design development to support site appraisal, brief preparation, concept development, spatial coordination, and stadium bay or stand optimisation based on quantifiable design characteristics. The framework addresses the inefficiencies and limitations of the traditional stadium design process by allowing rapid design space exploration defined by typological drivers, evaluation of a large set of solutions based on performance metrics such as circulation distances, sightline quality, and layout distribution, and the validation of concepts against benchmarks. Within the applicable design pipelines, and where labels are derived from deterministic performance criteria, the supervised approaches achieved prediction accuracies above 85%, while evolutionary optimisation reduced the number of seats with restricted views by approximately 95%. The value of the study is that it demonstrates that the integration of parametric modelling based on shared typological characteristics and the mapping of ML methods to the RIBA PoW has the potential to support stadium design in a novel way. Full article

High-precision topography of tidal flats is essential for coastal monitoring, geomorphic change analysis, and ecological assessment. Although satellite remote sensing supports repeated and large-area observation, topographic inversion over small-scale tidal flats—here defined as localized intertidal patches with limited areal extent, represented in this study by a 1.11 km² tidal flat near Dafeng Port—remains challenging, because ICESat-2 laser altimetry tracks across such areas are typically sparse and spatially clustered within narrow sub-regions, leaving extensive observation-blind zones without direct elevation labels. This label-clustering problem constrains the applicability of traditional empirical models and tends to cause deep learning models to generalize poorly beyond the spatial distribution of training samples. To address this issue, this study proposes a Residual Attention Physical-constraint Semi-supervised U-Net (RAPS-UNet) that fuses ICESat-2 ATL03/ATL08 elevation labels with Sentinel-1 SAR and Sentinel-2 optical features. The preprocessing pipeline comprises refined ICESat-2 photon filtering, adaptive inundation-frequency extraction, multi-source feature selection, and baseline DEM construction. RAPS-UNet integrates residual learning, attention-based multi-source fusion, physics-constrained loss, and confidence-weighted pseudo-label augmentation to improve extrapolation under clustered-label conditions. A four-level validation protocol—in-distribution validation, spatial holdout testing, and field-based assessment over both interpolation and extrapolation zones—was designed to evaluate spatial generalization. Against a field-surveyed DEM, RAPS-UNet achieved an overall RMSE of 0.20 m, an MAE of 0.16 m, and an R² of 0.91; the field-based interpolation and extrapolation zones yielded RMSEs of 0.17 m and 0.22 m, respectively, while the spatial holdout test reached an RMSE of 0.23 m and an R² of 0.81. Relative to the traditional inundation frequency–elevation linear model (RMSE = 0.35 m), RAPS-UNet reduced the field-validation RMSE by approximately 43%. The proposed framework therefore offers a practical approach for fine-scale coastal-zone topographic mapping under sparse and spatially clustered altimetry conditions. Full article