Earticle

The purpose of this study is to analyze and systematically classify the types of regular customers and the purpose of visiting fruit cafes through social big data analysis using Textom. As competition in the café market intensifies, the importance of customer retention strategies is increasing, and customers' visit experiences and opinions expressed on social media are becoming important marketing resources. In this study, 15,247 posts containing keywords such as 'fruit café', 'regular', and 'visit' were collected on SNS from January 2024 to December 2024 and text mining and semantic network analysis were performed. As a result of the analysis, regular customers of the fruit café were categorized into five types: 'health-seeking', 'atmosphere-oriented', 'workspace-oriented', 'communication-oriented', and 'taste exploration', and the purpose of visiting was categorized as 'maintaining healthy eating habits', 'escaping and healing from everyday life', 'utilizing work and study space', 'social interaction', and 'experiencing new tastes'. This study provides practical implications for establishing café marketing strategies and improving services by understanding café customers' needs and behavior patterns in depth through social big data analysis.

In this paper, we present LogiIQ, a VR-based logistics training platform designed to support repetitive learning and self-directed training to address the growing demand for skilled logistics personnel due to increasing industrial automation. We built LogiIQ on a Unity-based WebGL 3D environment to simulate the operation of key logistics equipment—robotic arms, AGVs, and forklifts. Furthermore, we implemented a system for continuous tracking of individual progress based on learning history by integrating a React and Spring Boot-based web dashboard to collect and visualize learner’s performance data in real time. In a pilot study involving novice users, we confirmed that average scores across all modules nearly doubled while the standard deviation decreased, thereby demonstrating that LogiIQ is effective in enhancing learner proficiency and reducing skill gaps among learners. By organically combining simulation, data analytics, and adaptive feedback, we have established a self-directed, data-driven learning environment that goes beyond simple virtual training. Through this research, we aim to lay a foundation for more consistent and effective practical training in logistics education.

In this paper, we propose the design and development of an AI-driven language learning application that leverages YouTube videos to provide immersive, personalized practice environments. The system empowers learners to choose any speaker from a YouTube video, then automatically identifies, isolates, and archives segments in which that individual appears. Once captured, these audio segments are converted into written transcripts, which serve as the foundation for interactive exercises including targeted listening comprehension, oral practice, and pronunciation refinement. A distinctive capability of the platform involves pausing playback precisely when the selected speaker talks, allowing learners to step into that character's role and participate in conversational exchanges with other figures in the video, thus strengthening their communicative abilities. Additionally, a specialized speaker-matching analysis module has been integrated to measure the acoustic similarity between learner output and the target speaker's voice. Through the extraction and comparison of acoustic characteristics—including vocal quality, articulation, tone, and melodic patterns—the system generates a numerical similarity rating on a 100-point scale. This comprehensive approach not only enhances multimodal interaction within second language learning environments but also establishes innovative pathways for tailored, immersive education grounded in genuine multimedia materials.

Accurate segmentation of breast tumors is essential for early cancer diagnosis, particularly in young women. Ultrasound imaging provides a non-invasive and cost-effective screening tool, yet breast ultrasound (BUS) segmentation is challenging due to tumor variability, blurred boundaries, and limited annotated datasets. To address annotation scarcity, we propose a semi-supervised framework that leverages both labeled and unlabeled BUS images. Pseudo-labels are generated for unlabeled data and refined with a multi-view strategy and self-consistency loss to improve reliability and stability. On the BUSI dataset, our method achieved Dice scores of 81%, 78%, and 76% with 1/2, 1/4, and 1/8 labeled data, respectively. We find these results to be equal to or superior to state-of-the-art approaches. Through this work, we can significantly reduce the annotation demands and expert labor required from radiologists.

Post-Training Quantization (PTQ), a method for model quantization without the need for retraining, is under active investigation for the deployment of real-time human pose estimation techniques on resourceconstrained edge devices. However, conventional PTQ techniques exhibit a critical limitation when applied to the SimCC output format utilized by the RTMPose model, leading to severe performance degradation due to the high sensitivity of its coordinate representation. This study proposes an optimization framework to overcome this challenge by systematically identifying optimal quantization targets through a Singular Value Decomposition (SVD)-based sensitivity analysis and by correcting the distorted output distribution via realtime post-processing. Experimental results demonstrate that the proposed framework achieves a 1.56-fold model compression while successfully retaining 88.8% of the accuracy of the original FP32 model. This research is anticipated to significantly enhance the practical deployment of models with sensitive output architectures, such as RTMPose, thereby contributing to a wide range of applications in real-time human pose estimation.

We purposed to enhance the accessibility and accuracy of alopecia counseling by proposing a retrievalaugmented generation (RAG) chatbot built on a large language model (LLM). The system integrates curated domain knowledge and empathy-oriented prompts to generate reliable and patient-friendly responses. A domain-specific dataset of approximately 1.33 million words was organized hierarchically and indexed with relevant keywords to support retrieval precision. To evaluate performance, 100 representative queries across five counseling categories were answered by both a baseline LLM and the proposed RAG-LLM. An external domain expert rated each response on factual accuracy, hallucination, clarity, and coverage. The RAG-LLM achieved higher factual accuracy (95% vs. 70%), lower hallucination (3% vs. 25%), and improved clarity and coverage (4.5 vs. 3.3). Statistical analysis confirmed the significance of these differences. The results demonstrate that platform-native RAG implementation can improve factual reliability and communication quality without external retrieval infrastructure, providing a practical framework for developing trustworthy, empathetic, and accessible digital health counseling systems.

Given the growing importance of online stores in sustaining consumer engagement, intellectual property (IP) character store designs increasingly use varied interactive features that jointly influence consumer behavior. We develop a model grounded in stimulus–organism–response theory to examine how an IP character store’s visual design, visibility, and social interaction affect user flow and engagement, which in turn shape consumer stickiness. Data were collected through an online survey targeting users who had purchased IP characters from a popular branded online store. We reveal that social interaction and visual design significantly enhance flow and engagement, thereby strengthening stickiness. These findings offer actionable insights for IP character management and online retail strategy.

Deep neural networks (DNNs) are widely deployed across safety-critical applications but remain vulnerable to adversarial attacks. This vulnerability is especially severe for compact models, which lack sufficient capacity to learn robust decision boundaries. While adversarial distillation (AD) offers a promising solution by transferring robustness from a large, robust teacher to a lightweight student, a fundamental limitation persists: as training progresses, the student generates increasingly strong adversarial examples that diverge from the teacher’s original adversarial domain. Since the teacher is typically fixed in conventional AD frameworks, its predictions become progressively less reliable on student-generated adversarial inputs, resulting in degraded supervision and limited robustness gains. To overcome this limitation, we propose ProxyAD, which equips the frozen teacher with a lightweight ProxyNet to dynamically align supervision with the student’s evolving adversarial inputs. ProxyNet is trained on student-generated adversarial examples using the teacher’s prediction on the corresponding unperturbed input as supervision, while a studentalignment regularizer prevents overconfident proxy soft labels on adversarial inputs by pulling them toward the student’s distribution so that targets remain calibrated and learnable during AD. This resulting supervision improves the student’s robustness with minimal additional parameters and modest training cost. Extensive experiments on CIFAR10/100 and Tiny-ImageNet show that ProxyAD consistently outperforms existing AD methods in both clean and adversarial accuracy under various attack scenarios.

Improved internet connectivity has accelerated the adoption of streaming video, with YouTube now hosting over 2 billion users worldwide. This study examines how influencer characteristics and content attributes shape Vietnamese travelers’ perceptions of destinations through YouTube. Data were collected from YouTube users in Vietnam via an online survey. The findings reveal that influencer characteristics and content attributes - including popularity, transparency, audience engagement, informativeness, real-time updates, visual appeal, and emotional connection - positively influence destination perceptions among Vietnamese travelers. This research offers both theoretical and practical implications for digital tourism marketing and influencer-driven content strategies.

This study aimed to analyze research trends in depression among Korean university students over the past decade using a big data–driven topic modeling approach. To conduct the analysis, 803 articles published from 2015 to 2024 were retrieved from three major Korean academic databases, namely DBpia, KCI, and RISS. After data preprocessing and synonym/stopword adjustments, 1,971 valid keywords were extracted. Latent Dirichlet Allocation (LDA) was performed using NetMiner 4.5.1 to identify latent topics and temporal changes. The analysis identified three major topics: (1) Psychosocial Stress and Emotional Well-being, focusing on stress, anxiety, self-esteem, and interpersonal relationships; (2) Health and Pandemic-related Issues in Nursing Students, highlighting mental health concerns of nursing students, particularly during COVID-19; and (3) Social Support, Life Satisfaction, and Digital Influence, emphasizing social connectedness, quality of life, and the impact of smartphone addiction. The distribution of topics revealed that pandemic-related research (Topic-2) increased significantly after 2020, while social support and digital environment studies (Topic-3) relatively declined. These findings demonstrate that university student depression research in Korea has evolved around psychosocial, pandemic-related, and social/digital dimensions. The study highlights the academic value of topic modeling in mapping research landscapes and suggests the need for multidimensional interventions, ranging from individual stress management to structural support for specific student subgroups.

In this study, to accurately estimate targets in a maritime environment, the radar cross section (RCS) is modeled using a Weibull distribution, and likelihood maximization together with QR decomposition is applied to the RML algorithm to minimize the angle-of-arrival (AoA) error. In the simulation experiments conducted with a 64-element antenna array and a Signal-to-Noise Ratio (SNR) of 23 dB, the proposed algorithm demonstrated significant performance improvement compared to the conventional RML algorithm. Specifically, the estimation error rate for the Angle of Arrival (AOA) was reduced from 0.058 degrees (conventional RML) to 0.04 degrees (proposed method), which represents a 31% reduction in directional error. Furthermore, the proposed algorithm substantially reduced the velocity estimation error during the initial target extraction phase, showing an estimated reduction of approximately 4 m/s compared to the RML method. The simulation also clarified the experimental setup by conducting 1,000 Monte Carlo runs to ensure reproducibility. The results statistically validate that applying QR decomposition to RML offers superior performance in maritime target extraction by effectively removing surface-signal clutter.

In this paper, the calibration path implementation of the Ku band tracking antenna system with about 2000 arrangements and the verification method of the calibration function during the near field test were described. The calibration path of the tracking antenna system is an internal loop consisting of an antenna element assembly, a TR assembly, an RF power combiner assembly, a driving amplifier module, and an RF cable. The gain and phase change for each channel for calibration are performed by the 4CH Beamformer IC inside the TR assembly. After extracting the align data and calibration data, during the near field test process, the change in the antenna system state was imitated by changing the RF cable inside the antenna system. The CLUT(Calibration Look Up Table) was calculated by re-extracting the antenna system calibration data after the state change and calculating the difference from the antenna system calibration data before the state change. Thereafter, by applying the alignment LUT(Look Up Table) loaded into the antenna system before the state change and the CLUT through the scenario of the RF test equipment, it was verified whether the antenna system maintains the alignment state, the calibration function, and the effectiveness of the calibration value when the state of the antenna system changes.

This study proposes a vehicle positioning system that integrates RTK-based Global Navigation Satellite System (GNSS), an Inertial Measurement Unit (IMU), and wheel-tick data obtained from the vehicle’s On-Board Diagnostics (OBD) to overcome the frequent GNSS signal outages and multipath effects that occur in urban, underground, and tunnel environments. To ensure stable positioning even under unstable RTK signal conditions, an Extended Kalman Filter (EKF) was applied. Experimental results show that the proposed system improves positioning accuracy by approximately 5 meters in open-sky environments using RTK. Under Dead Reckoning (DR) conditions, integrating OBD-based wheel-tick data with the IMU reduced the position error by up to 30 meters compared to using the IMU alone. Furthermore, in tunnel-driving experiments, the system maintained a position error within 3.5 meters even when GNSS signals were completely lost. The proposed method demonstrates practical applicability in various domains requiring high-precision positioning, such as autonomous driving, high-definition map construction, and real-time vehicle monitoring.

Through color, children express their emotions, show creativity, and are motivated to learn. However, currently, children’s color education is not systematic in many schools, which reduces the consistency and effectiveness of color education. As a result, children are limited in their opportunities to experience and utilize various colors. In this study, a children's color program was developed to get children to develop and understand the properties of colors, as well as to experience the five senses, and express them with 700 color chips, and to improve their emotions, creativity, and imagination based on this. This study emphasizes the importance of children's color education and proposes effective educational methods and programs. Findings of the study are as follows: It was found that the various feelings of observation of taste, smell, hearing, and touch, including sight, were meaningful. Above all, it was meaningful in that creative thinking was possible by breaking away from the method of simply viewing and drawing the subject and spreading the imagination through storytelling techniques. It is hoped that this study will increase interest in children's color education as well as contribute to the development of practical children's color education programs.

Bubble defects reduce the reliability of PCB coatings. Accurate, real-time detection of small bubbles is crucial. In this study, we present a fast and lightweight YOLOv8n-based detector. It improves the average precision of small bubbles by approximately 12 points. It also increases mAP50 from 0.637 to 0.697 in 49.1 milliseconds per image. This is a practical solution that supports quality control of micro-defects. NAMAttention reweights channel and spatial features. C2F fuses backbone and neck features. The P2 layer expands the receptive field for micro-bubbles. During training, size-aware loss emphasizes small bubbles. Defect-balanced sampling addresses class imbalance. Glare synthesis improves robustness to illumination variations. During inference, Soft-NMS reduces false positives caused by overlapping boxes. Our contribution lies in the organic integration of a learning strategy specialized for micro-defects and architectural improvements. At the same time, we maintain a lightweight structure and improve detection performance and latency. We quantitatively verified the utility and practical feasibility of each component through module-bymodule erasure experiments. We also performed throughput evaluations in field-deployment scenarios. This approach is effective for PCB quality control and suitable for detecting micro-defects in similar manufacturing processes.

A novel broadband and wide-angle polarization beam-splitter (PBS) based on a dumbbell grating with a cylindrical profile is proposed and numerically analyzed in this study. The designed structure consists of alternating Si and ZnS cylindrical resonators stacked on a SiO2 substrate. By employing the Eigenvalue Problem of Modal Transmission-Line Theory (EP-MTLT), the reflection and transmission characteristics for TE and TM polarizations are rigorously investigated. The cylindrical dumbbell profile enhances polarizationdependent guided-mode resonance (GMR), providing strong reflection for TE-polarized light and efficient transmission for TM-polarized light across a broad wavelength range. Simulation results demonstrate that the proposed PBS achieves high polarization extinction ratio, broadband operation, and wide angular tolerance up to ±20°, while maintaining compact geometry and fabrication robustness. Therefore, the proposed structure offers a promising approach for realizing miniaturized, broadband, and wide-angle polarization beamsplitters suitable for photonic integration and optical communication systems.

We explore the potential for blockchain technology and the Web3.0 paradigm to innovate the complex intellectual property (IP) trading ecosystem. Focusing on structural paradoxes within the creator economy and the rise of derivative works, this paper introduces the concept of a User-Defined IP Ecosystem, envisioning a new environment where users take the lead, enabling transparent and streamlined transactions. We employ an interdisciplinary approach consisting of three stages: First, we apply Transaction Cost Economics to analyze how blockchain reduces search, negotiation, and enforcement costs. Second, we utilize Network Effects theory to examine value amplification mechanisms within derivative ecosystems. Third, we conduct architectural analysis of the STORY platform to demonstrate practical implementation. Through this framework, the paper examines the mechanics of a virtuous cycle where blockchain's cost-reduction benefits and network effects' value amplification reinforce each other. This analysis reveals new possibilities for innovation in traditional IP business models and illustrates the reconfiguration of the web ecosystem, where users regain agency in transactions. Ultimately, this study suggests that the User-Defined IP Ecosystem innovation could be a critical turning point, redefining user experience, key player roles, and interactions within the digital asset trading ecosystem.

In this study, according to the design requirements of smoke exhaust systems in Chinese and Korean building codes, a typical university laboratory building was selected as a target, and a three-dimensional fire model was set up with PyroSim software.Article 51, Paragraph 2, Item 1 of the Enforcement Decree of the Korean Building Act: Article 14, Article 14, 1 Color, 2 Colors Asterice 2 (Flue Window Standards) of the Regulations on Building Equipment Standards for the Installation of 6 or More Types of Smoke Emission Facilities and the difference in smoke emission efficiency between the Chinese GB51251 standards was compared and analyzed. The simulation results showed that the larger smoke emission area (3.2% vs. 2.5%) required by the Korean standard extends the delay time of flue gas layer descent by 48 seconds and reduces the maximum CO concentration by 37%. We impact to an optimal basis for the architectural design of laboratory buildings in multinational universities.

This study conducted a comparative analysis of the fire development characteristics under different combustible materials and national smoke exhaust window standards, with a focus on the regulations for university laboratory buildings in China and Korea. Using the PyroSim fire simulation software, a full-scale three-dimensional model was established based on the typical architectural form of the laboratory, and according to the surface parameters of the smoke exhaust Windows in China and Korea. To comprehensively assess the fire dynamics, three representative fuel sources - ethanol, paper and wood - were selected to reflect various laboratory fire scenarios. Under natural exhaust conditions, key parameters such as temperature distribution, smoke visibility, oxygen consumption, concentrations of carbon monoxide (CO) and carbon dioxide (CO2) were analyzed. The research results emphasize that the design parameters of smoke exhaust Windows and the characteristics of combustants jointly affect the fire safety outcome. We impact to this research provides valuable data for optimizing the fire safety design of laboratories, informing about the revision of regulations, and promoting cross-border fire engineering research.

The Fourth Industrial Revolution and the proliferation of artificial intelligence (AI) are transforming the structure of education, and EduTech has become a key area of educational innovation. This study aims to analyze unstructured data from 2020 to 2025 to identify the time-series changes in AI-based EduTech and suggest a structural direction for the Artificial Intelligence-based Digital Transformation for Education (AIDX). LDA, BERTopic topic modeling, and social network analysis (SNA) were applied to identify relationships between key topics and keywords. The analysis revealed that EduTech has evolved from its initial focus on "AI learning systems" to encompass "generative AI," "personalized learning," and "AI ethics," evolving into an intelligent ecosystem that combines technology, industry, and policy. These results demonstrate that AIDX is moving beyond simple technology adoption and is acting as a structural driver of educational innovation.

This study proposes an AI-based wireless network temperature measurement system designed to capture and monitor fine-grained temperature variations within a grid-partitioned indoor environment. Unlike conventional thermostats that provide only a few point measurements, the proposed system employs a dense array of wireless sensor nodes distributed across a spatial grid to achieve high-resolution spatial coverage. Sensor readings are transmitted in real time to a centralized data aggregation server, which performs data cleaning, anomaly filtering, and integration with historical records. The core of the approach is an LSTMbased prediction model capable of short-horizon forecasting of temperature dynamics in each grid cell by leveraging temporal dependencies across multi-sensor data streams. By combining dense sensing, robust preprocessing, and predictive analytics, the system enhances spatial resolution and improves both the accuracy and responsiveness of indoor climate control. Early detection of anomalies such as localized hotspots or cooling inefficiencies allows proactive HVAC adjustments that reduce energy consumption, mitigate thermal discomfort, and prevent operational faults. Field experiments demonstrate that the proposed framework delivers real-time forecasting with low latency and high detection sensitivity, supporting adaptive decision-making in dynamic indoor environments. This integrated sensing-and-prediction paradigm represents a scalable and energy-aware solution for next-generation smart buildings, contributing to improved occupant comfort, sustainable energy management, and overall operational efficiency.

In this paper, we present a novel AI-assisted IPTV broadcasting system (AIBS) for real-time automatic VOD subtitle generation. The proposed system generates subtitles using Digital TV-Closed Caption (DTV-CC) data and decides the display time of VOD subtitles via utilizing the latest AI technology. With this approach, the problems of subtitle accuracy and display time errors of existing STT systems can be resolved. First, we introduce a method for obtaining subtitle information through DTV-CC in the existing IPTV broadcasting system. Then, the align function of the Stable-TS is adopted to obtain the display time of the subtitles and, for a more accurate time, various noises other than voices in the input audio signal are filtered with deep learningbased background noise removal technology as pre-processing. Finally, if the subtitle display time provided by Stable-TS is abnormal, the display time information is corrected using an NLP-based post-processing technology that utilizes the speaker separation information and relative time information of DTV-CC. Experimental results shows that the average accuracy, which was 87% when using Stable-TS alone, improved by 7% to an average of 94% after applying AIBS on testing the accuracy of subtitle synchronization for approximately 3,000 sentences. In addition, by applying AIBS to the existing IPTV broadcasting system that generates quick-VOD based on linear channels, the problem of manually generating VOD subtitles that took 4 hours was reduced to less than 10 minutes, improving customer satisfaction.

The advent of quantum computing raises concerns about data security, a core feature of blockchain technology. This study refines and extends existing analyses by separating the encapsulation mechanism (KEM) and the digital signature algorithm (DSA), thereby conceptualizing correctness. Using the Open Quantum Safe (OQS) library, this study benchmarked Kyber (KEM), Dilithium, FALCON, and SPHINCS+ (DSA), quantifying on-chain impacts such as signature size, gas costs, and verification latency on an Ethereumcompatible testnet. The results of this study demonstrate that Dilithium offers a balanced tradeoff between security and throughput, while SPHINCS+ remains impractical in high-transaction-per-second environments. Kyber demonstrates superior KEM efficiency, making it suitable for session key distribution in hybrid blockchain architectures.