Document classification is a complicated task in the early stages for enterprises or companies. In the past, documents had to be read by humans and sent to the departments by ourselves; this also costs a lot of human resources and may cause errors because there are too many departments; nowadays, document classification turns more autonomously and given up the traditional way of processing documents, many places setting up systems to process the documents even uses the technology of OCR and LLM to implement into the system to help perform classifying tasks to the assigned department and send it automatically. In this research, we will experiment with LLM models to observe which is more suitable for document classification after fine-tuning. We will utilize BERT (Bidirectional Encoder Representations from Transformers) and RoBERTa (Robustly optimized BERT approach) in the experiment, as they are designed to classify and understand Chinese texts, embedding them into our document classification system to help direct documents to the correct assigned department.

Traffic sign datasets often suffer from data scarcity and class imbalance, which challenge the development of robust autonomous driving systems. This article proposes a novel dataset augmentation method that leverages Stable Diffusion inpainting to generate realistic synthetic traffic signs. The method fine-tunes a Stable Diffusion model and introduces an object-size-based crop (OSB-crop) technique with mask adjustments to ensure high-quality augmentations that maintain contextual consistency. Evaluations using the Fréchet Inception Distance (FID) show average scores of 195.85 for the DFG-T10 subset and 247.077 for the DFG-B10 subset, demonstrating the ability to produce realistic inpainted signs, particularly for more represented minority classes. Qualitative analyses further highlight seamless integration into real-world scenes, although challenges remain for extremely underrepresented classes and ensuring perfect visual fidelity. The benefits of this approach include its potential to enhance traffic sign datasets, address class imbalances, and improve the potential for training more reliable autonomous driving systems by providing more diverse and realistic training data. This study focuses on evaluating the quality of the generated data itself as a foundational step toward enhancing downstream detection models. However, limitations include the computational cost of fine-tuning and the difficulty in achieving high-quality inpainting for all underrepresented classes, especially those with poor initial data quality. This work lays a strong foundation for advancing dataset augmentation techniques for real-world applications.

Image processing effects are powerful tools for data augmentation in deep learning, potentially expanding the diversity of training data for object detection. However, while techniques like dehazing aim to restore visibility, applying them without careful consideration can introduce artifacts or cause domain shifts, harming model performance. In this research, we explore the dual nature of dehazing—as both a helpful enhancer and a potential distorter—specifically within the context of the YOLOv12 architecture. We propose a rigorous methodology to determine the optimal dehaze intensity for training by establishing a theoretical upper bound via brute-force search. Additionally, we introduce a feature-consistency analysis that sorts augmented samples into four categories based on their effect on the backbone’s feature representation. Our experiments reveal a key link between model capacity and the effectiveness of augmentation. We observe that implicit feature enhancement using “Neutral” (Tier 0) samples yields significant improvements of 0.84 mAP_50-95 for the YOLOv12-Small and YOLOv12-Large models, respectively. However, it negatively affects YOLOv12-Nano, resulting in a -0.9

We present LiteYOLOv8, a lightweight pedestrian detection architecture optimized for real-time edge deployment. Built upon YOLOv8n, it integrates a FasterNet backbone for efficient feature extraction, a BiFPN for multi-scale fusion, and SimAM attention to refine spatial focus. On the CityPersons dataset, LiteYOLOv8 improves mAP@50 by 2.2%, reduces parameters by 65%, and lowers FLOPs by 21%. Real-device evaluation on a Raspberry Pi 5 confirms low latency and minimal resource usage, demonstrating its suitability for autonomous and surveillance applications.

Pedestrian safety at urban intersections remains a major challenge in Intelligent Transportation Systems (ITSs). This study investigates whether crossing intention can be reliably inferred directly from temporal body-pose dynamics to drive real-time collision warnings on embedded edge platforms. Existing vision-based approaches that rely primarily on bounding-box proximity or scene-level spatial grids are often prone to false alarms in complex urban environments with motorcycles, stationary pedestrians, and background clutter. To overcome these limitations, we propose an end-to-end framework consisting of four sequential processing stages: (1) a perception layer integrating YOLOv8s, ByteTrack, a displacement filter, and rider suppression to generate reliable pedestrian trajectories; (2) a skeleton extraction layer utilizing YOLOv8s-pose to construct temporal sequences of 17 anatomical keypoints; (3) an ultra-lightweight Skeleton Graph Convolutional Network (SkeletonGCN, comprising 33.8 K parameters, <0.2 MB) that models body-joint kinematics and temporal motion dynamics; and (4) an image-space Time-to-Collision (TTC) risk-fusion module. While this fusion approach avoids explicit geometric camera calibration, it still relies on predefined scene-profile parameters and image-space motion assumptions. Furthermore, while the intention classifier is quantitatively evaluated, the risk-fusion module is procedurally defined, and its resulting four-level collision warnings are demonstrated operationally rather than validated against ground-truth hazard annotations. Evaluated on 49,948 valid sequences from the JAAD and PIE benchmark datasets under a strict video-level partitioning protocol, the unified SkeletonGCN achieves a macro-F1 score of 0.717 (with per-scene subset macro-F1 scores of 0.761 on JAAD/PIE urban and 0.895 on intersections), significantly outperforming baseline models. When deployed on an NVIDIA Jetson Orin NX edge device using TensorRT FP16, the full pipeline achieves an instrumented latency of 70.7 ms per frame (~14 fps) and a sustained wall-clock throughput of 7.4 fps on real-world urban dashcam video. System limitations include sensitivity to 2D printed human imagery and reduced prediction reliability under low-light nighttime conditions.

This study investigates the use of Vision-Language Models (VLMs), particularly YOLO-World, for industrial anomaly detection, emphasizing the need for explicit defect localization through bounding boxes. We adapt the MVTec AD dataset by converting segmentation masks to bounding boxes and employing a rigorous 1:9 few-shot split from the test set. By comparing a Zero-Shot approach (text prompts only) with a Few-Shot Fine-Tuning method (using fewer than 15 images per class), we find that Zero-Shot inference struggles with abstract defects (mAP@50<2%). In contrast, our specialized few-shot fine-tuning-freezing the backbone and training only the head-yields significant improvements, such as a 56.5% mAP@50 for the Metal Nut category with about 10 training samples. This highlights the potential of VLMs for low-data industrial inspection.

Urban traffic congestion is increasingly caused by the inefficiency of locating available parking spaces. This study proposes a vision-based segmentation approach for detecting real-time parking availability using YOLOv8 on a custom dataset collected from dashcam footage in urban Taiwan. The dataset comprises over 200 carefully annotated images, augmented to 513 samples, and segmented into parking-area and no-parking-area classes. A YOLOv8 instance segmentation model was trained to detect and localize open parking spaces with pixel-level precision. The model achieved a mask of 0.983 and mask :0.95 of 0.784, demonstrating strong performance in complex, real-world road scenarios. These results highlight the potential of deep learning-based segmentation for smart parking applications in dense urban environments, offering a scalable, accurate, and sensor-free alternative suitable for real-time deployment.

Automated surface defect detection often faces a "cold-start" problem due to limited annotated data for new anomalies. Traditional object detectors struggle to converge in such few-shot settings. To address this, we adapt Vision-Language Models (VLMs), specifically YOLO-World. We use semantic pre-training to mitigate data scarcity. We evaluate this approach on the MVTec AD dataset in bounding-box format. We use a strict 1:9 train-validation split, resulting in an average of 11.8 defect instances per category. YOLO-World surpasses traditional baselines, like YOLOv11s and YOLOv26s, in 12 of 15 categories. The optimized VLM pipeline achieves up to 64.9% mAP@50 on texture-heavy categories, such as Tile, with only nine training instances. Ablation studies show standard optimization techniques are limited under 10-shot constraints. We find a critical augmentation divide. Disabling spatial distortions (Mosaic) is vital to preserving rigid-object geometry. The Normalized Wasserstein Distance (NWD) improves the localization of microscopic anomalies. Varifocal Loss (VFL) often causes model collapse. Ultimately, VLMs offer a superior foundation for cold-start inspection but require carefully tailored pipelines for robustness.

Driver adaptation between left-side and right-side traffic systems often involves delayed reactions and unsafe visual checks. Prior research on driver visual attention typically requires specialized eye tracking hardware or immersive VR systems, limiting cost effective deployment. This paper proposes a desktop driving simulation framework that integrates webcam-based gaze estimation to quantify visual attention, mirror checking accuracy, reaction time, and lane control. A controlled pre/post training design evaluates both left to right and right to left transitions. Paired t-tests and repeatedmeasures ANOVA reveal significant post training improvements in all primary metrics. The results support the utility of webcam-based gaze analysis for assessing driver adaptation without expensive sensors.

Pedestrian safety at urban intersections is a critical issue in Intelligent Transportation Systems (ITS). Traditional safety analyses largely rely on reactive accident statistics, failing to capture dynamic human-vehicle interactions in real-time. This paper proposes an end-to-end traffic safety warning framework leveraging raw surveillance video. The system architecture, as illustrated in Fig. 1, integrates three core modules: (1) A Perception Layer utilizing a fine-tuned YOLOv8m with ByteTrack, enhanced by novel Rider Suppression and Displacement Filter mechanisms to mitigate noise; (2) A Semantic Understanding Layer employing hybrid maps (vehicle trajectories and visual features) to define navigable areas; and (3) A Risk Prediction Layer based on Spatio-Temporal Graph Convolutional Networks (ST-GCN). By discretizing intersections into spatio-temporal grids, the model simulates time-varying collision risks. Experimental results on real-world datasets demonstrate a Precision of 85.7% and Recall of 89.4%, validating the system's effectiveness in proactive risk assessment.

Rapid urban expansion and economic activity in densely populated cities have significantly contributed to worsening air quality. Jakarta, Indonesia’s capital, frequently experiences PM2.5 concentrations that exceed "unhealthy" levels, posing serious public health risks. While various studies have explored air pollution modeling, there remains a gap in integrating localized educational and ethical frameworks into predictive approaches. This study applies the Fuzzy k-Nearest Neighbor (FKNN) algorithm to classify PM2.5 concentration levels in Jakarta using meteorological variables as predictive inputs. With a data split of 80

On-street parking significantly contributes to urban traffic congestion and emissions. This research addresses the challenging problem of real-time detection of unmarked parallel parking spaces, which lack clear visual demarcations and exhibit variable appearances. We present an optimized YOLOv5s model that overcomes the limitations of traditional convolutional approaches by incorporating a C3TR module (Transformer-based) for capturing long-range spatial dependencies and a PANet head for robust multi-scale feature aggregation. These enhancements effectively enable the model to handle occlusions, varying lighting conditions, and diverse parking space sizes. The model was trained and evaluated on a comprehensive dataset of marked and unmarked parallel parking scenarios. 5-fold cross-validation confirms the model's robustness, consistency, and generalizability, achieving an average mAP@50 of 80.3%. The system's computational efficiency makes it suitable for deployment on resource-constrained devices, offering a practical solution for improving parking management and reducing urban congestion.

This research implements the YOLOv10 model on the UP7000 embedded system for real-time road marking detection and evaluates two lightweight variants: YOLOv10n and YOLOv10s. YOLOv10s achieved higher precision (P), F1-score, and mAP50, indicating improved accuracy and reduced false positives, while YOLOv10n demonstrated higher recall (R), ensuring a greater detection rate of road markings. In a real-world deployment, YOLOv10n exhibited better inference speed, reaching 9 FPS, whereas YOLOv10s achieved only 4 FPS, suggesting that YOLOv10n is more suitable for high-speed traffic scenarios. The system detects 10 types of road markings and traffic lights, automatically recording GPS coordinates, timestamps, and images for further analysis. This scalable solution enhances urban infrastructure management and is suitable for smart city applications.

Road marking signs are part of traffic signs on the road's surface. It can be found on almost every section of the road. It gives information about the road lanes and helps the driver understand the road ahead. Located on the road surface, road marking signs are more vulnerable to damage. Maintaining the condition of road marking signs is very challenging. It is a combination of environmental, logistical, and financial factors. Keeping the road marking signs clear and visible is essential for road safety. Faded or damaged road marking signs are difficult for the driver to see and identify, which can lead to missing or wrong information being captured. This research proposes a novel method to detect damaged road marking signs and classify the damage level using the combination of YOLOv9, VAE, image processing techniques, and fuzzy logic. The experiment results show that YOLOv9-M performs best in general road marking sign detection with 0.92 precision, 0.906 recall, 0.935 mAP@50, and 0.668 mAP@50:95. YOLOv9-M also excels in detecting the damaged road marking signs with 0.872 detection rate. The combination of VAE and image processing techniques shows promising results in detecting damaged road marking signs with 0.8395 precision, 0.8212 recall, and 0.7295 accuracy. The combination of anomaly score from the VAE with fuzzy logic is used to classify the damage level into three categories, which is more meaningful in understanding the road marking sign condition.

Text summarization plays an important role in learning enhancement by extracting valuable insights from the original text. From an educational point of view, an effective dialogue summary can help students, teachers, and researchers structure the content better by getting precise and concise content. This research investigates the use of a fine-tuned T5-base model for creating abstractive summaries of educational dialogues, utilizing the SAMSum dataset, which consists of dialogue-based interactions accompanied by human-crafted summaries. This research adds to educational technologies by introducing a summarization model that enhances reading comprehension in academic settings, particularly for learners with extensive digital materials or those facing learning difficulties. We assessed the model's efficiency through the ROUGE metric, obtaining ROUGE-1: 0.50, ROUGE-2: 0.25, and ROUGE-L: 0.41, showcasing competitive results relative to current state-of-the-art models. The experimental findings showed that our model performed better.

Child malnutrition continues to be a significant public health issue in Indonesia, with considerable spatial variations among provinces. This study utilizes an integrated framework that combines Bayesian Network (BN) modeling, Local Indicators of Spatial Association (LISA), and Moran’s I analysis to ascertain the principal determinants of child nutritional status and their spatial distribution. The Bayesian Network (BN) probabilistic approach captures intricate interdependencies among health and behavioral factors, such as Correct Knowledge about Stunting (CKS), Exclusive Breastfeeding (EBF), Breastfeeding for Two Years (BF2Y), Complementary Feeding (CF), Immunization (IMM), Growth Monitoring (GM), Iron-Folic Acid Supplementation (IFA), Antenatal Care (ANC6), and Animal Protein Intake for Pregnant Women (APH), and their impact on outcomes classified as Severely Underweight (SUW), Underweight (UW), and Normal (N). Spatial statistical analyses indicate considerable heterogeneity and elevated-risk clusters of malnutrition, underscoring areas necessitating prioritized interventions. Stochastic simulations elucidate the variability and robustness of critical determinants, whereas model evaluation via Mean Squared Error (MSE) indicates robust predictive performance for variables such as N, ANC6, and SUW, with heightened uncertainty noted for EBF, IFA, and IMM, implying the impact of supplementary social or contextual factors. This study combines probabilistic inference with spatial analysis to show the pathways and interactions that lead to child malnutrition. This gives useful information for focused, evidence-based interventions. The results support strategies that involve more than one sector and are in line with Indonesia’s National Strategy for Stunting Reduction (Presidential Regulation No. 72 of 2021). This will help resources be used more effectively, policies be planned better, and children’s nutrition outcomes be more fair across the country.

Robust visual perception and geometric alignment are crucial for intelligent automation in various domains, such as industrial processes and infrastructure monitoring. Accurately aligning structured visual elements, such as floor markings or road-marking templates, is essential for tasks like automated guidance, verification, and condition assessment. However, traditional feature-based methods struggle with templates that feature simple geometries and lack rich textures, making reliable feature matching and alignment difficult, even under controlled conditions. To address this, we propose GeoTemplateKPNet, a novel self-supervised deep-learning framework, built upon Convolutional Neural Networks (CNNs), designed to learn robust, geometrically consistent keypoints specifically in synthetic template images. The model is trained exclusively in a synthetic template dataset by enforcing equivariance to geometric transformations and utilizing self-supervised losses, including inside mask loss, peakiness loss, repulsion loss, and keypoint-driven image reprojection loss, thereby eliminating the need for manual keypoint annotations. We evaluate the method in a synthetic template test set, using metrics such as a keypoint-matching comparison, the Inside Mask Rate (IMR), and the Alignment Reconstruction Error (ARE). The results demonstrate that GeoTemplateKPNet successfully learns to predict meaningful keypoints on template structures, enabling accurate alignment between templates and their transformed counterparts. Ablation studies reveal that the number of keypoints (K) impacts the performance, with K = 3 providing the most suitable balance for the overall alignment accuracy, although the performance varies across different template geometries. GeoTemplateKPNet offers a foundational self-supervised solution for the robust geometric analysis of templates, which is crucial for downstream alignment tasks and applications.

Unlike the detection of marked on-street parking spaces, detecting unmarked spaces poses significant challenges due to the absence of clear physical demarcation and uneven gaps caused by irregular parking. In urban cities with heavy traffic flow, these challenges can result in traffic disruptions, rear-end collisions, sideswipes, and congestion as drivers struggle to make decisions. We propose a real-time detection system for on-street parking spaces using YOLO models and recommend the most suitable space based on KD-tree search. Lightweight versions of YOLOv5, YOLOv7-tiny, and YOLOv8 with different architectures are trained. Among the models, YOLOv5s with SPPF at the backbone achieved an F1-score of 0.89, which was selected for validation using k-fold cross-validation on our dataset. The Low variance and standard deviation recorded across folds indicate the model’s generalizability, reliability, and stability. Inference with KD-tree using predictions from the YOLO models recorded FPS of 37.9 for YOLOv5, 67.2 for YOLOv7-tiny, and 67.0 for YOLOv8. The models successfully detect both marked and unmarked empty parking spaces on test data with varying inference speeds and FPS. These models can be efficiently deployed for real-time applications due to their high FPS, inference speed, and lightweight nature. In comparison with other state-of-the-art models, our models outperform them, further demonstrating their effectiveness.

This study investigates the efficient retrieval of traffic accident locations near vehicle travel routes, a critical component for real-time navigation systems aimed at enhancing driver safety. Utilizing Taiwan's 2023 traffic accident dataset, which serves as historical data for identifying accident locations along routes in OpenStreetMap, we compare the performance of four spatial search methods: Buffer Analysis, Nearest Neighbor Analysis, KD-tree, and R-tree. These methods are evaluated based on construction time, query time, total computation time, and maximum memory usage across various route lengths and accident densities, including four tourist destinations in Taiwan. Our results demonstrate that the KD-tree algorithm significantly outperforms the others in query performance and total computation time, making it the best choice for real-time accident alerts in navigation systems. This study provides an effective solution for traffic accident data analysis and route planning to improve driver safety.