Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Haiyoung Jung has a total of 35 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Scenario-Wise Comparative Evaluation of Lightweight YOLO Models for Flame and Smoke Detection

    by Sugi Choi, SaengGon Lee, Hyunjeong Bang, YoungJoo Song, Haiyoung Jung

    Flames and smoke vary in appearance across fire environments, limiting the ability of aggregate detection metrics to characterize model applicability. This study compared lightweight models from ten YOLO versions using an integrated dataset comprising five fire scenarios. Basic image statistics and DINOv2 feature embeddings characterized the scenario groups, while class- and scenario-specific evaluation examined performance differences obscured by aggregate metrics. All models used the same fixed data splits. Aggregate test results were summarized as means and sample standard deviations across three training runs using seeds 0, 21, and 42. The evaluation covered detection accuracy, object-size-specific average precision, inference speed, and fixed-threshold error characteristics. No single model achieved the highest class-specific AP50-95 across all scenarios. Wildland and outdoor fires yielded AP50-95 values of 0.285–0.358 for flame and 0.293–0.392 for smoke, whereas smoke AP50-95 exceeded 0.90 for all evaluated models in electrical equipment and industrial facility fires and in road and vehicle fires. Small objects also remained challenging across models. These findings demonstrate how scenario-stratified evaluation reveals detection weaknesses beyond aggregate rankings and supports baseline selection according to the target environment, object class, and computational requirements.

  • Open Access

    A Study on GIS-Based Regional Analysis of Relative Safety Vulnerability Using Nationwide Road Bridge Data

    by Jiwon Choi, Haiyoung Jung

    This study presents a data-driven screening framework for comparing the relative safety vulnerability of road bridges and identifying bridges requiring priority review before detailed fire-safety assessment. We created a nationwide dataset comprising 40,006 road bridges in South Korea in 2025; among these; 33,466 bridges with valid information on the location, traffic volume, completion year, and design load were selected for analysis. The traffic volume, years in service, and design load were normalized using min‑max scaling and integrated into a relative safety vulnerability index (R) using exploratory weights of 0.5, 0.3, and 0.2, respectively. These variables represented use intensity, aging, and structural capacity instead of the fire occurrence probability, fire load, or structural fire resistance. The resulting index was classified into five levels, compared across 17 administrative regions, and spatially visualized using maps based on the Geographic Information System. Among the 33,466 bridges analyzed, 14,471 were classified as high or very-high vulnerability, accounting for 43.2% of the valid analysis dataset. In terms of the proportion of high-vulnerability bridges relative to the total number of registered bridges in each region, Gyeongsangnam-do (38.4%), Chungcheongnam-do (38.2%), and Jeollanam-do (38.1%) had the highest values, whereas Daejeon Metropolitan City had the lowest (24.9%). The proposed index is intended as a preliminary screening and prioritization tool for selecting bridges that warrant field inspection and subsequent fire-safety assessment and not as a direct estimate of bridge fire risk.

  • Open Access

    AI-Generated Fire Images for Object Detection-Based Fire Detection

    by Wangeun Ji, Sugi Choi, Heejun Kwon, Haiyoung Jung

    Vision-based fire detection models are often limited by the insufficient diversity of annotated fire and smoke images, particularly in terms of fire location, flame scale, smoke density, ignition cause, and indoor scene context. This study investigates whether generative AI-based synthetic images can expand fire-image diversity and improve object detection-based fire detection performance. Real fire images were combined with conventional augmented images and synthetic images generated using ChatGPT-4.o and ChatGPT-5.5. The generated images were constructed using multivariable prompts considering fire location, scale, and cause, and unsuitable samples were screened using a pretrained fire detection model. YOLOv8n, YOLOv11n, and RT-DETR were trained under 48 dataset–detector conditions and evaluated using fixed validation and test datasets. The results showed that generated-image-based training generally maintained or improved detection performance compared with the original and conventional augmentation conditions. In particular, selected ChatGPT-4.o-based YOLOv11 conditions showed statistically supported improvements over matched augmentation conditions, with increases of +0.052 in Precision, +0.031 in Recall, +0.065 in mAP@0.5, and +0.038 in mAP@0.5:0.95. LPIPS and t-SNE analyses indicated that the generated images formed structured perceptual and feature-space distributions relative to real fire images. Scenario-based inference using location-specific video frames also showed stable model responses in several complex indoor fire environments. These findings suggest that validated generative AI-based images can supplement the limited visual diversity of real fire datasets and improve the robustness of vision-based fire detection models.

  • Open Access

    Analysis of the Effect of Generative AI-Based Data Augmentation on Early Wildfire Smoke Detection Performance

    by Sangwon Lee, Haiyoung Jung

    This study analyzes the effect of generative artificial intelligence (AI)-based data augmentation on early wildfire smoke detection performance. Early-stage wildfire smoke is difficult to detect because of its low density, poor contrast against the background, and indistinct boundaries. To address these limitations, early wildfire smoke images were selected from the AI-Hub wildfire dataset, and augmented datasets were constructed using conventional image augmentation, contrastive unpaired translation (CUT)-based image generation, and diffusion-based image generation using Stable Diffusion XL (SDXL) with low-rank adaptation (LoRA). The augmented images were only added to the training set, and the You Only Look Once version 11 nano (YOLOv11n) model was used to evaluate detection performance under the same validation and test conditions. Precision, recall, F1-score, mAP_0.5, and mAP_0.5:0.95 were used as evaluation metrics. The results demonstrated that SDXL LoRA-based diffusion augmentation with 500 generated images achieved the highest performance, with an F1-score of 0.786 and an mAP_0.5:0.95 of 0.520. In particular, mAP_0.5:0.95 increased from 0.499 under the original condition to 0.520, indicating that using an appropriate number of diffusion-generated images can improve the localization performance of early wildfire smoke detection. However, diffusion augmentation with 1,000 generated images increased recall while decreasing precision, implying that excessive synthetic data can result in false positives. Therefore, generative AI-based data augmentation for early wildfire smoke detection should consider not only the number of synthetic images but also their visual consistency and detection efficacy.

  • Open Access

    A Study on Performance Analysis of Electric Vehicle Fire and Smoke Object Detection Based on Different Labeling Methods Using the YOLOv11 Model

    by Jiwon Choi, Eunji Yoon, Haiyoung Jung

    This study evaluated the performance of YOLOv11 models trained to detect electric vehicle (EV) fires and smoke emission using bounding-box and instance segmentation annotations. Comprehensive experiments were conducted across all YOLOv11 variants (n, s, m, l, x) to analyze the trade-offs between speed and detection accuracy. The models were trained on a dataset consisting of 3,000 images depicting EV fires and smoke, which were annotated using Roboflow’s polygon tool under identical training conditions with consistent hardware and hyperparameters. The evaluation was based on precision, recall, and mean average precision at an intersection over union threshold of 0.50 (mAP50), as well as mAP50-95, F1-score, and inference speed in frames per second. The experimental results showed distinct advantages for each annotation method. Bounding-box models demonstrated superior inference speed, with the YOLOv11n model achieving the highest speed at 44.62 FPS. This makes these methods optimal for real-time EV fire monitoring applications. By contrast, segmentation models showed significantly higher accuracy in detecting objects with irregular boundaries, particularly smoke plumes. The YOLOv11m model with segmentation achieved optimal performance with a mAP50 of 0.7952 and an F1-score of 0.7510. The segmentation approach proved particularly effective for smoke detection, where bounding-box annotation frequently failed to accurately capture the irregular boundaries of smoke in an image, which resulted in detection failures for these critical fire indicators. These findings demonstrate that segmentation labeling is more effective in achieving accurate early detection and precise localization of EV fires and smoke emissions, despite the computational trade-off. The superior boundary definition capability of the segmentation approach is crucial for reliable smoke detection, which provides an important early warning of thermal runaway events in EV batteries. Future studies will focus on expanding the diversity of the dataset used to train the models by applying generative AI techniques such as GAN and diffusion models to further enhance detection robustness across varied fire scenarios.

  • Open Access

    Machine Learning-Based On-Device Smoke and Flame Detection System Using Light Intensity of RGB Color Sensor

    by Minsung Park, Bo-Hee Lee, Haiyoung Jung

    This study proposes an edge-based real-time fire-detection system utilizing low-cost RGB sensors and TinyML technology. To overcome the high cost and maintenance limitations of existing detection methods in large indoor spaces, we propose an efficient approach that analyzes light-intensity patterns using affordable RGB sensors. Unlike simple threshold-based methods, our system applies machine learning to precisely distinguish between the complex optical patterns of smoke and flames, thus significantly enhancing detection accuracy. Designed with an on-device AI architecture, the system performs independent and rapid fire assessment on an edge device without requiring external network connectivity. Experiments conducted in a simulated fire environment confirmed the system’s feasibility and high accuracy. Spectral analysis and multisensor fusion shall be considered in future investigations to further minimize false alarms.

  • Open Access

    Dynamic Behavior of Electric Arc in Low-Voltage Series Arc Faults on the Basis of Load Power Factor Characteristics

    by Sunghwan Kim, Haiyoung Jung

    This study investigates the effect of the load power factor (PF) on the electrical arc characteristics of low-voltage series arcs. The arc voltage and current waveforms are measured under high- and low-power-factor load conditions, with 100 sweeps averaged to derive the conductance G(t) and its time derivative dG(t)/dt. Based on these data, key parameters of the Cassie–Mayr hybrid model—including the Cassie voltage (Uc), Mayr power (Pm), arc time constant (θh), and transition conductance (Gt)—are estimated using nonlinear least squares optimization. The analysis reveals that in the arc generation phase, low-power-factor loads are associated with higher values of Pm and Gt, indicating more explosive arc initiation with more rapid energy release when compared to those observed under stable and gradual heat accumulation and high-power-factor conditions. Conversely, during the arc extinction phase, high-power-factor loads are associated with a better Pm and longer θh, reflecting stronger resistance to arc extinction and a tendency to sustain heat over time. Overall, these results provide essential data for the quantitative assessment of electrical arc risks, and they serve as a foundation for the development of arc fault detection and interruption technologies.

  • Open Access

    Comparative Assessment of YOLO Segmentation Extensions for Intelligent Fire Detection

    by Sugi Choi, Heejun Kwon, Jiwon Choi, Sangwon Lee, Haiyoung Jung

    With the growing frequency of fire incidents, the demand for rapid and accurate fire detection technologies has become increasingly critical. In this study, we evaluate segmentation-based object detection models YOLO (You Only Look Once) v5-seg, YOLOv8-seg, and YOLOv11-seg for their ability to detect flames and smoke under identical experimental conditions. A total of 5,000 fire images were collected and split into training, validation, and test datasets. The same hardware environment and hyperparameter settings were used for model training to ensure a fair comparison. The experimental results reveal that YOLOv11-seg achieved the best overall performance, with a Precision of 0.710, Recall of 0.570, F1-score of 0.632, and mAP (mean Average Precision) 50 of 0.600. Notably, YOLOv11-seg achieved the highest Recall and mAP values for smoke detection, underscoring its effectiveness in identifying smoke—a critical factor for early fire detection. In terms of efficiency, YOLOv8-seg demonstrated the fastest inference speed, while YOLOv5-seg offered advantages in lightweight model size. However, YOLOv11-seg provided a balanced trade-off between computational cost and detection accuracy, making it the most suitable model for real-world fire response scenarios. Accordingly, this study proposes YOLOv11-seg as a robust baseline model for segmentation-based fire detection and provides a foundational reference for future research on deep learning-driven intelligent fire video analysis.

  • Open Access

    Evaluation of Satisfaction and Practical Effectiveness of GIS-Based Firefighting Training Systems among Firefighters

    by Hyunjung Bang, Sugi Choi, Hyunjin Nam, Seunghee Oh, Haiyoung Jung

    The existing mobile data terminal (MDT)-based dispatch systems are limited in their ability to reflect detailed spatial information, such as floor plans, multiple access points, and the locations of hazardous materials. As a complementary solution, geographic information system (GIS)-based firefighting training systems have attracted interest. This study evaluated the field applicability and effectiveness of a GIS-based training platform by conducting a scenario-based fire response drill at the OO Industrial Complex Fire Station in Incheon, South Korea. A post-training survey was administered to 30 participating firefighters. The analysis revealed high levels of satisfaction in terms of dispatch response, including MDT-based dispatch mapping (80%), situational information provided by the control center (76%), and access to information on fire targets (70%). For training effectiveness, 83% of respondents found the GIS data useful, and 83% rated the system as more effective than the existing method. Moreover, 73% agreed that the integration of Global Navigation Satellite System (GNSS) and GIS data significantly contributed to directional suppression and resource allocation decisions. Age-based analysis showed generally high satisfaction across all age groups, with particularly strong acceptance among firefighters in their 30s or younger, whereas a more conservative attitude was observed among those aged 40 and above. These findings empirically demonstrate that GIS-based firefighting training systems effectively overcome the limitations of conventional tabletop training and significantly enhance both dispatch efficiency and strategic response capabilities.

  • Open Access

    Optimized Faster R-CNN with Swintransformer for Robust Multi-Class Wildfire Detection

    by Sugi Choi, Sunghwan Kim, Haiyoung Jung

    Wildfires are a critical global threat, emphasizing the need for efficient detection systems capable of identifying fires and distinguishing fire-related from non-fire events in their early stages. This study integrates the swintransformer into the Faster R-CNN backbone to overcome challenges in detecting small flames and smoke and distinguishing complex scenarios like fog/haze and chimney smoke. The proposed model was evaluated using a dataset comprising five classes: flames, smoke, clouds, fog/haze, and chimney smoke. Experimental results demonstrate that swintransformer-based models outperform ResNet-based Faster R-CNN models, achieving a maximum mAP50 of 0.841 with the swintransformer-based model. The model exhibited superior performance in detecting small and dynamic objects while reducing misclassification rates between similar classes, such as smoke and chimney smoke. Precision–recall analysis further validated the model’s robustness across diverse scenarios. However, slightly lower recall for specific classes and a lower FPS compared to ResNet models suggest a need for further optimization for real-time applications. This study highlights the swintransformer’s potential to enhance wildfire detection systems by addressing fire and non-fire events effectively. Future research will focus on optimizing its real-time performance and improving its recall for challenging scenarios, thereby contributing to the development of robust and reliable wildfire detection systems.