Search Results (103)

Semantic segmentation of remote sensing images is crucial for geospatial applications but is severely hampered by the prohibitive cost of pixel-level annotations. Although semi-supervised learning (SSL) offers a solution by leveraging unlabeled data, its application to Vision Transformers (ViTs) often encounters overfitting and even training instability under extreme label scarcity. To tackle these challenges, we propose a Curriculum-based Self-supervised and Semi-supervised Pipeline (CSSP). The pipeline adopts a staged, easy-to-hard training strategy, commencing with in-domain pretraining for robust feature representation, followed by a carefully designed finetuning stage to prevent overfitting. The pipeline further integrates a novel Difficulty-Adaptive ClassMix (DA-ClassMix) augmentation that dynamically reinforces underperforming categories and a Progressive Intensity Adaptation (PIA) strategy that systematically escalates augmentation strength to maximize model generalization. Extensive evaluations on the Potsdam, Vaihingen, and Inria datasets demonstrate state-of-the-art performance. Notably, with only 1/32 of the labeled data on the Potsdam dataset, the CSSP reaches 82.16% mIoU, nearly matching the fully supervised result (82.24%). Furthermore, we extend the CSSP to a semi-supervised domain adaptation (SSDA) scenario, termed Cross-Domain CSSP (CDCSSP), which outperforms existing SSDA and unsupervised domain adaptation (UDA) methods. This work establishes a stable and highly effective framework for training ViT-based segmentation models with minimal annotation overhead. Full article

For the problem of online real-time prediction of copper particle content in the lubricating oil of the main spindle-bearing system of mining equipment, the traditional direct detection method is costly and has insufficient real-time performance. To this end, this paper proposes an indirect prediction method based on data-driven neural networks. The proposal of this method is based on a core assumption: during the stable wear stage of the equipment, there exists a modelable statistical correlation between the copper particle content in the oil and the total amount of non-ferromagnetic particles that are easy to measure online. Based on this, a neural network prediction model was constructed, with the online metal abrasive particle sensor signal (non-ferromagnetic particle content) as the input and the copper particle content as the output. The experimental data are derived from 100 real oil samples collected on-site from the lubrication system of the main shaft bearing of a certain mine mill. To enhance the model’s performance in the case of small samples, data augmentation techniques were adopted in the study. The verification results show that the average prediction accuracy of the proposed neural network model reaches 95.66%, the coefficient of determination (R²) is 0.91, and the average absolute error (MAE) is 0.3398. Its performance is significantly superior to that of the linear regression model used as the benchmark (with an average accuracy of approximately 80%, R² = 0.71, and the mean absolute error (MAE) = 1.5628). This comparison result not only preliminarily verified the validity of the relevant hypotheses of non-ferromagnetic particles and copper particles in specific scenarios, but also revealed the nonlinear nature of the relationship between them. This research explores and preliminarily validates a low-cost technical path for the online prediction of copper particle content in the stable wear stage of the main shaft bearing system, suggesting its potential for engineering application within specific, well-defined scenarios. Full article

Although general-domain Named Entity Recognition (NER) has achieved substantial progress in the past decade, its application to cybersecurity NER is hindered by the lack of publicly available annotated datasets, primarily because of the sensitive and privacy-related nature of security data. Prior research has largely sought to improve performance by expanding annotation volumes, while overlooking the intrinsic characteristics of training data. In this study, we propose a cybersecurity Named Entity Recognition (NER) method based on hard and easy labeled training data discrimination. Firstly, a hybrid strategy that integrates a deep learning (DL)-based discriminator and a rule-based discriminator is employed to partition the original dataset into hard and easy samples. Secondly, the proportion of hard and easy data in the training set is adjusted to determine the optimal balance. Finally, a data augmentation algorithm is applied to the partitioned dataset to further improve model performance. The results demonstrate that, under a fixed total training data size, the ratio of hard to easy samples has a significant impact on NER performance, with the optimal strategy achieved at a 1:1 proportion. Moreover, the proposed method further improves the overall performance of cybersecurity NER. Full article

Background/Objectives: Traditional dietary monitoring methods such as 24 h recalls rely on self-report, leading to recall bias and underreporting. Similarly, dietary control approaches, including portion control and calorie restriction, depend on user accuracy and consistency. Augmented reality (AR) offers a promising alternative for improving dietary monitoring and control by enhancing engagement, feedback accuracy, and user learning. This systematic review aimed to examine how AR technologies are implemented to support dietary monitoring and control and to evaluate their usability and effectiveness among adults. Methods: A systematic search of PubMed, CINAHL, and Embase identified studies published between 2000 and 2025 that evaluated augmented reality for dietary monitoring and control among adults. Eligible studies included peer-reviewed and gray literature in English. Data extraction focused on study design, AR system type, usability, and effectiveness outcomes. Risk of bias was assessed using the Cochrane RoB 2 tool for randomized controlled trials and ROBINS-I for non-randomized studies. Results: Thirteen studies met inclusion criteria. Since the evidence based was heterogeneous in design, outcomes, and measurement, findings were synthesized qualitatively rather than pooled. Most studies utilized smartphone-based AR systems for portion size estimation, nutrition education, and behavior modification. Usability and satisfaction varied by study: One study found that 80% of participants (N = 15) were satisfied or extremely satisfied with the AR tool. Another reported that 100% of users (N = 26) rated the app easy to use, and a separate study observed a 72.5% agreement rate on ease of use among participants (N = 40). Several studies also examined portion size estimation, with one reporting a 12.2% improvement in estimation accuracy and another showing −6% estimation, though a 12.7% overestimation in energy intake persisted. Additional outcomes related to behavior, dietary knowledge, and physiological or psychological effects were also identified across the review. Common limitations included difficulty aligning markers, overestimation of amorphous foods, and short intervention durations. Despite these promising findings, the existing evidence is limited by small sample sizes, heterogeneity in intervention and device design, short study durations, and variability in usability and accuracy measures. The limitations of this review warrant cautious interpretation of findings. Conclusions: AR technologies show promise for improving dietary monitoring and control by enhancing accuracy, engagement, and behavior change. Future research should focus on longitudinal designs, diverse populations, and integration with multimodal sensors and artificial intelligence. Full article

This paper presents a markerless augmented reality (AR) navigation system for guiding users across a university campus, independent of internet or wireless connectivity, integrating machine learning (ML) and deep learning techniques. The system employs computer vision to detect campus signage “Meeting Point” and “Directory”, and classifies them through a binary classifier (BC) and convolutional neural networks (CNNs). The BC distinguishes between the two types of signs using RGB values with algorithms such as Perceptron, Bayesian classification, and k-Nearest Neighbors (KNN), while the CNN identifies the specific sign ID to link it to a campus location. Navigation routes are generated with the Floyd–Warshall algorithm, which computes the shortest path between nodes on a digital campus map. Directional arrows are then overlaid in AR on the user’s device via ARCore, updated every 200 milliseconds using sensor data and direction vectors. The prototype, developed in Android Studio, achieved over 99.5% accuracy with CNNs and 100% accuracy with the BC, even when signs were worn or partially occluded. A usability study with 27 participants showed that 85.2% successfully reached their destinations, with more than half rating the system as easy or very easy to use. Users also expressed strong interest in extending the application to other environments, such as shopping malls or airports. Overall, the solution is lightweight, scalable, and sustainable, requiring no additional infrastructure beyond existing campus signage. Full article

Purpose: While hybrid quantum–classical neural networks (HNNs) are a promising avenue for quantum advantage, the critical influence of the classical backbone architecture on their performance remains poorly understood. This study investigates the role of lightweight convolutional neural network architectures, focusing on LCNet, in determining the stability, generalization, and effectiveness of hybrid models augmented with quantum layers for medical applications. The objective is to clarify the architectural compatibility between quantum and classical components and provide guidelines for backbone selection in hybrid designs. Methods: We constructed HNNs by integrating a four-qubit quantum circuit (with trainable rotations) into scaled versions of LCNet (050, 075, 100, 150, 200). These models were rigorously evaluated on CIFAR-10 and MedMNIST using stratified 5-fold cross-validation, assessing accuracy, AUC, and robustness metrics. Performance was assessed with accuracy, macro- and micro-averaged area under the ROC curve (AUC), per-class accuracy, and out-of-fold (OoF) predictions to ensure unbiased generalization. In addition, training dynamics, confusion matrices, and performance stability across folds were analyzed to capture both predictive accuracy and robustness. Results: The experiments revealed a strong dependence of hybrid network performance on both backbone architecture and model scale. Across all tests, LCNet-based hybrids achieved the most consistent benefits, particularly at compact and medium configurations. From LCNet050 to LCNet100, hybrid models maintained high macro-AUC values exceeding 0.95 and delivered higher mean accuracies with lower variance across folds, confirming enhanced stability and generalization through quantum integration. On the DermaMNIST dataset, these hybrids achieved accuracy gains of up to seven percentage points and improved AUC by more than three points, demonstrating their robustness in imbalanced medical settings. However, as backbone complexity increased (LCNet150 and LCNet200), the classical architectures regained superiority, indicating that the advantages of quantum layers diminish with scale. The mostconsistent gains were observed at smaller and medium LCNet scales, where hybridization improved accuracy and stability across folds. This divergence indicates that hybrid networks do not necessarily follow the “bigger is better” paradigm of classical deep learning. Per-class analysis further showed that hybrids improved recognition in challenging categories, narrowing the gap between easy and difficult classes. Conclusions: The study demonstrates that the performance and stability of hybrid quantum–classical neural networks are fundamentally determined by the characteristics of their classical backbones. Across extensive experiments on CIFAR-10 and DermaMNIST, LCNet-based hybrids consistently outperformed or matched their classical counterparts at smaller and medium scales, achieving higher accuracy and AUC along with notably reduced variability across folds. These improvements highlight the role of quantum layers as implicit regularizers that enhance learning stability and generalization—particularly in data-limited or imbalanced medical settings. However, the observed benefits diminished with increasing backbone complexity, as larger classical models regained superiority in both accuracy and convergence reliability. This indicates that hybrid architectures do not follow the conventional “larger-is-better” paradigm of classical deep learning. Overall, the results establish that architectural compatibility and model scale are decisive factors for effective quantum–classical integration. Lightweight backbones such as LCNet offer a robust foundation for realizing the advantages of hybridization in practical, resource-constrained medical applications, paving the way for future studies on scalable, hardware-efficient, and clinically reliable hybrid neural networks. Full article

Making time-critical decisions with serious consequences is a daily aspect of work environments. To support the process of finding optimal actions, data-driven approaches are increasingly being used. The most advanced form of data-driven analytics is prescriptive analytics, which prescribes actionable recommendations for users. However, the produced recommendations rely on complex models and optimization techniques that are difficult to understand or justify to non-expert users. Currently, there is a lack of platforms that offer easy integration of domain-specific prescriptive analytics workflows into production environments. In particular, there is no centralized environment and standardized approach for implementing such prescriptive workflows. To address these challenges, large language models (LLMs) can be leveraged to improve interpretability by translating complex recommendations into clear, context-specific explanations, enabling non-experts to grasp the rationale behind the suggested actions. Nevertheless, we acknowledge the inherent black-box nature of LLMs, which may introduce limitations in transparency. To mitigate these limitations and to provide interpretable recommendations based on real user knowledge, a knowledge graph is integrated. In this paper, we present and validate a prescriptive analytics platform that integrates ontology-based graph retrieval-augmented generation (GraphRAG) to enhance decision making by delivering actionable and context-aware recommendations. For this purpose, a knowledge graph is created through a fully automated workflow based on an ontology, which serves as the backbone of the prescriptive platform. Data sources for the knowledge graph are standardized and classified according to the ontology by employing a zero-shot classifier. For user-friendly presentation, we critically examine the usability of GraphRAG in prescriptive analytics platforms. We validate our prescriptive platform in a customer clinic with industry experts in our IoT-Factory, a dedicated research environment. Full article

In order to ensure the earthquake safety of existing buildings, retrofitting applications come to the fore in terms of being fast and cost-effective. Among these applications, fiber-reinforced polymer (FRP) composites are widely preferred thanks to their advantages such as high strength, corrosion resistance, applicability without changing the cross-section and easy assembly. This study presents a data augmentation, modeling, and comparison-based approach to predict the fire resistance (FR) of FRP-strengthened reinforced concrete beams. The aim of this study was to explore the role of data augmentation in enhancing prediction accuracy and to find out which augmentation method provides the best prediction performance. The study utilizes an experimental dataset taken from the existing literature. The dataset contains inputs such as varying geometric dimensions and FRP-strengthening levels. Since the original dataset used in the study consisted of 49 rows, the data size was increased using augmentation methods to enhance accuracy in model training. In this study, Gaussian noise, Regression Mixup, SMOGN, Residual-based, Polynomial + Noise, PCA-based, Adversarial-like, Quantile-based, Feature Mixup, and Conditional Sampling data augmentation methods were applied to the original dataset. Using each of them, individual augmented datasets were generated. Each augmented dataset was firstly trained using eXtreme Gradient Boosting (XGBoost) with 10-fold cross-validation. After selecting the best-performing augmentation method (Adversarial-like) based on XGBoost results, the best-performing augmented dataset was later evaluated in HyperNetExplorer, a more advanced NAS tool that can find the best performing hyperparameter optimized ANN for the dataset. ANNs achieving R² = 0.99, MSE = 22.6 on the holdout set were discovered in this stage. This whole process is unique for the FR prediction of structural elements in terms of the data augmentation and training pipeline introduced in this study. Full article

Large language models (LLMs) now match or exceed human performance on many open-ended language tasks, yet they continue to produce fluent but incorrect statements, which is a failure mode widely referred to as hallucination. In low-stakes settings this may be tolerable; in regulated or safety-critical domains such as financial services, compliance review, and client decision support, it is not. Motivated by these realities, we develop an integrated mitigation framework that layers complementary controls rather than relying on any single technique. The framework combines structured prompt design, retrieval-augmented generation (RAG) with verifiable evidence sources, and targeted fine-tuning aligned with domain truth constraints. Our interest in this problem is practical. Individual mitigation techniques have matured quickly, yet teams deploying LLMs in production routinely report difficulty stitching them together in a coherent, maintainable pipeline. Decisions about when to ground a response in retrieved data, when to escalate uncertainty, how to capture provenance, and how to evaluate fidelity are often made ad hoc. Drawing on experience from financial technology implementations, where even rare hallucinations can carry material cost, regulatory exposure, or loss of customer trust, we aim to provide clearer guidance in the form of an easy-to-follow tutorial. This paper makes four contributions. First, we introduce a three-layer reference architecture that organizes mitigation activities across input governance, evidence-grounded generation, and post-response verification. Second, we describe a lightweight supervisory agent that manages uncertainty signals and triggers escalation (to humans, alternate models, or constrained workflows) when confidence falls below policy thresholds. Third, we analyze common but under-addressed security surfaces relevant to hallucination mitigation, including prompt injection, retrieval poisoning, and policy evasion attacks. Finally, we outline an implementation playbook for production deployment, including evaluation metrics, operational trade-offs, and lessons learned from early financial-services pilots. Full article

Cyanotoxins in agricultural waters pose a human and animal health risk. These toxins can be transported to nearby crops and soil during irrigation practices; they can remain in the soil for extended periods and be adsorbed by root systems. Additionally, in livestock watering ponds, cyanotoxins pose a direct ingestion risk. This work evaluated the performance of the random forest algorithm in estimating microcystin concentrations using eight in situ water quality measurements at one active livestock water pond and two working irrigation ponds in Georgia and Maryland, USA. Measurements of microcystin along with eight in situ-sensed water quality parameters were used to train and test the machine learning model. The models performed better at the Georgia ponds compared to the Maryland pond, and interior models performed better than nearshore or whole-pond models. The most important variables for microcystin prediction were water temperature and phytoplankton pigments. Overall, the random forest algorithm(RF), augmented with a ‘trainControl’ function to perform repeated cross validations, was able to explain 40% to 70% of the microcystin concentration variation in the three agricultural ponds. Water quality measurements showed potential to aid water monitoring/sampling design by predicting the microcystin concentrations in the studied ponds by using readily available and easy to collect in situ data. Full article

This research presents the development of a type-2 fuzzy-controlled autonomous mobile robot specifically designed for monitoring and actively maintaining indoor air quality. The core of this system is the proposed type-2 fuzzy PID dual-mode controller used for stably patrolling rooms along the walls of the environment. The design method ingeniously merges the fast error correction capability of PID control with the robust adaptability of type-2 fuzzy logic control, which utilizes interval type-2 fuzzy sets. Furthermore, the type-2 fuzzy rule table of the right wall-following controller can be extended from the first designed fuzzy left wall-following controller in a symmetrical design manner. As a result, this study eliminates the drawbacks of excessive oscillations arising from PID control and sluggish response to large initial errors in typical traditional fuzzy control. The following of the stable wall and obstacle is facilitated with ensured accuracy and easy implementation so that effective air quality monitoring and active PM2.5 filtering are achieved in a movable manner. Furthermore, the augmented reality (AR) interface overlays real-time PM2.5 data directly onto a user’s visual field, enhancing situational awareness and enabling an immediate and intuitive assessment of air quality. As this type of control is different from that used in traditional fixed sensor networks, both broader area coverage and efficient air filtering are achieved. Finally, the experimental results demonstrate the controller’s superior performance and its potential to significantly improve indoor air quality. Full article

Sleep position recognition plays a crucial role in diagnosing and managing various health conditions, such as sleep apnea, pressure ulcers, and musculoskeletal disorders. Accurate monitoring of body posture during sleep can provide valuable insights for clinicians and support the development of intelligent healthcare systems. This research presents a comparative analysis of sleep position recognition using two distinct approaches: image-based deep learning and accelerometer-based classification. There are five classes: prone, supine, right side, left side, and wake up. For the image-based method, the Visual Geometry Group 16 (VGG16) convolutional neural network was fine-tuned with data augmentation strategies including rotation, reflection, scaling, and translation to enhance model generalization. The image-based model achieved an overall accuracy of 93.49%, with perfect precision and recall for “right side” and “wakeup” positions, but slightly lower performance for “left side” and “supine” classes. In contrast, the accelerometer-based method employed a feedforward neural network trained on features extracted from segmented accelerometer data, such as signal sum, standard deviation, maximum, and spike count. This method yielded superior performance, reaching an accuracy exceeding 99.8% across most sleep positions. The “wake up” position was particularly easy to detect due to the absence of body movements such as heartbeat or respiration when the person is no longer in bed. The results demonstrate that while image-based models are effective, accelerometer-based classification offers higher precision and robustness, particularly in real-time and privacy-sensitive scenarios. Further comparisons of the system characteristics, data size, and training time are also carried out to offer crucial insights for selecting the appropriate technology in clinical, in-home, or embedded healthcare monitoring applications. Full article

Nowadays, financial institutions increasingly leverage artificial intelligence to enhance decision-making and optimize investment strategies. A specific application is the automatic analysis of large volumes of unstructured textual data to extract relevant information through deep learning (DL) methods. However, the effectiveness of these methods is often limited by the scarcity of high-quality labeled data. To address this, we propose a new data augmentation technique, Concatenation Augmentation (CA). This is designed to overcome the challenges of processing unstructured text, particularly in analyzing professional profiles from corporate governance reports. Based on Mixup and Label Smoothing Regularization principles, CA generates new text samples by concatenating inputs and applying a convex additive operator, preserving its spatial and semantic coherence. Our proposal achieved hit rates between 92.4% and 99.7%, significantly outperforming other data augmentation techniques. CA improved the precision and robustness of the DL models used for extracting critical information from corporate reports. This technique offers easy integration into existing models and incurs low computational costs. Its efficiency facilitates rapid model adaptation to new data and enhances overall precision. Hence, CA would be a potential and valuable data augmentation tool for boosting DL model performance and efficiency in analyzing financial and governance textual data. Full article

The categories for call center conversation data are valuably used for reporting business results and marketing analysis. However, they typically lack clear patterns and suffer from severe imbalance in the number of instances across categories. The call center conversation categories used in this study are Payment, Exchange, Return, Delivery, Service, and After-sales service (AS), with a significant imbalance where Service accounts for 26% of the total data and AS only 2%. To address these challenges, this study proposes a model that ensembles meta-information generated through Named Entity Recognition (NER) with machine learning inference results. Utilizing KoBERT (Korean Bidirectional Encoder Representations from Transformers) as our base model, we employed Easy Data Augmentation (EDA) to augment data in categories with insufficient instances. Through the training of nine models, encompassing KoBERT category probability weights and a CatBoost (Categorical Boosting) model that ensembles meta-information derived from named entities, we ultimately improved the F1 score from the baseline of 0.9117 to 0.9331, demonstrating a solution that circumvents the need for expensive LLMs (Large Language Models) or high-performance GPUs (Graphic Process Units). This improvement is particularly significant considering that, when focusing solely on the category with a 2% data proportion, our model achieved an F1 score of 0.9509, representing a 4.6% increase over the baseline. Full article

Data analysis is widely used in research and industry where there is a need to extract information from data. A significant amount of time within a data analysis project is required to prepare the data for subsequent analysis. This paper presents a comprehensive weighted maturity model to estimate the readiness of data for subsequent data analysis, with the goal of avoiding delays due to data quality problems. The maturity model uses a questionnaire with nine criteria to determine the maturity level of data preparation. The maturity model is integrated into a web application that provides an automated evaluation of maturity and a novel visualization approach that combines a modified spider chart and augmented chord diagrams. The comprehensive weighted maturity model is a ready-to-use application that provides prospective users with an easy and quick way to check their data for maturity for subsequent data analysis, with the goal of improving the data preparation process. The weighted maturity model is applicable to all types of data analysis, regardless of the domain of the data. Full article