Sang-Hyun Lee

This study proposes a lightweight crack-segmentation model optimized for industrial and edge-computing environments, where both high accuracy and real-time inference are required. Conventional convolution-based and U-Net-based crack segmentation models offer relatively simple architectural designs, but often suffer from limited boundary precision or an unfavorable accuracy–efficiency trade-off. Swin Transformer-based approaches can model broader contextual information but may still show poor segmentation quality relative to their computational cost in fine crack analysis. To address these limitations, we propose the Stabilized Crack Network++ (SCN++), a U-Net backbone crack segmentation network that integrates edge fusion, hybrid loss with deep supervision, exponential moving average (EMA)-based stabilization, and lightweight post-processing. The model was trained and evaluated on 40,000 concrete surface images, including 20,000 crack images and 20,000 non-crack images, using quantitative metrics such as intersection over union (IoU), Dice coefficient, frames per second (FPS), giga floating-point operations (GFLOPs), and the number of parameters, together with overlay-based qualitative analysis. Compared with the CNN, U-Net, and Swin Transformer baselines, SCN++ achieved the best overall balance between segmentation accuracy and computational efficiency, with an IoU of 0.7346, a Dice coefficient of 0.8457, 35.09 FPS, 8.45 GFLOPs, and only 2.22 M parameters. These results demonstrate that SCN++ effectively mitigates the conventional accuracy–efficiency trade-off and is a strong candidate for practical structural crack segmentation in edge-computing environments.

Accurate and timely detection of rice leaf diseases is critical for ensuring global food security and maximizing agricultural yields. However, existing deep learning methods often struggle to balance the high accuracy required for detecting multi-scale lesions in complex field environments with the computational efficiency necessary for edge device deployment. This paper proposes You Only Look Once for Lightweight Detection (YOLOv11-LD), a lightweight object detection model for multi-scale rice leaf disease detection in real paddy field environments. The model is built on YOLOv11n and integrates a Re-parameterized Vision Transformer (RepViT) backbone, a Bidirectional Feature Pyramid Network (BiFPN) based neck, and a Convolutional Block Attention Module (CBAM) to enhance multi-scale feature representation to enhance multi-scale feature representation while maintaining a lightweight architecture suitable for edge deployment. A dataset of 3234 images captured in actual rice paddies was constructed, containing three major rice leaf diseases: bacterial blight, rice blast, and brown spot. and was split into 2241 training images and 993 validation images. Ablation experiments show that the full YOLOv11-LD configuration achieves 95.2% mAP_0.5 with 7.8 Giga Floating-Point Operations (GFLOPs) and 3.5M parameters, outperforming the baseline YOLOv11n (91.4% mAP_0.5) under the same input resolution of 640 × 640. Additional comparisons with Faster Region-based Convolutional Neural Network (Faster R-CNN), Single Shot MultiBox Detector (SSD), YOLOv5n, YOLOv8n, and YOLOv11n further confirm that YOLOv11-LD provides the best overall trade-off between detection accuracy and computational efficiency. These results demonstrate that YOLOv11-LD offers superior operational efficiency suitable for resource-constrained smart rice disease monitoring systems.

This study proposes a lightweight rice disease detection model optimized for edge computing environments. The goal is to enhance the You Only Look Once (YOLO) v5 architecture to achieve a balance between real-time diagnostic performance and computational efficiency. To this end, a total of 3234 high-resolution images (2400 × 1080) were collected from three major rice diseases Rice Blast, Bacterial Blight, and Brown Spot—frequently found in actual rice cultivation fields. These images served as the training dataset. The proposed YOLOv5-V2 model removes the Focus layer from the original YOLOv5s and integrates ShuffleNet V2 into the backbone, thereby resulting in both model compression and improved inference speed. Additionally, YOLOv5-P, based on PP-PicoDet, was configured as a comparative model to quantitatively evaluate performance. Experimental results demonstrated that YOLOv5-V2 achieved excellent detection performance, with an mAP 0.5 of 89.6%, mAP 0.5–0.95 of 66.7%, precision of 91.3%, and recall of 85.6%, while maintaining a lightweight model size of 6.45 MB. In contrast, YOLOv5-P exhibited a smaller model size of 4.03 MB, but showed lower performance with an mAP 0.5 of 70.3%, mAP 0.5–0.95 of 35.2%, precision of 62.3%, and recall of 74.1%. This study lays a technical foundation for the implementation of smart agriculture and real-time disease diagnosis systems by proposing a model that satisfies both accuracy and lightweight requirements.

This study proposes a deep learning-based long short-term memory (LSTM) model to predict the state of charge (SOC) of lithium-ion batteries. The purpose of the research is to accurately model the complex nonlinear behavior that occurs during the charging and discharging processes of batteries to predict the SOC. The LSTM model was trained using battery data collected under various temperature and load conditions. To evaluate the performance of the artificial intelligence model, measurement data from the CS2 lithium-ion battery provided by the University of Maryland College of Engineering was utilized. The LSTM model excels in learning long-term dependencies from sequence data, effectively modeling temporal patterns in battery data. The study trained the LSTM model based on battery data collected from various charge and discharge cycles and evaluated the model’s performance by epoch to determine the optimal configuration. The proposed model demonstrated high SOC estimation accuracy for various charging and discharging profiles. As training progressed, the model’s predictive performance improved, with the predicted SOC moving from 14.8400% at epoch 10 to 12.4968% at epoch 60, approaching the actual SOC value of 13.5441%. Simultaneously, the mean absolute error (MAE) and root mean squared error (RMSE) decreased from 0.9185% and 1.3009% at epoch 10 to 0.2333% and 0.5682% at epoch 60, respectively, indicating continuous improvement in predictive performance. In conclusion, this study demonstrates the effectiveness of the LSTM model for predicting the SOC of lithium-ion batteries and its potential to enhance the performance of battery management systems.

Modern aquaculture utilizes computer vision technology to analyze underwater images of fish, contributing to optimized water quality and improved production efficiency. The purpose of this study is to efficiently perform underwater fish detection and tracking using multi-object tracking (MOT) technology. To achieve this, the FairMOT model was employed to simultaneously implement pixel-level object detection and re-identification (Re-ID) functions, comparing two backbone models: FairMOT+YOLOv5s and FairMOT+DLA-34. The study constructed a dataset targeting the popular black porgy in Korean aquaculture, using underwater video data from five different environments collected from the internet. During the training process, the FairMOT+YOLOv5s model rapidly reduced train loss and demonstrated stable performance. The FairMOT+DLA-34 model showed better results in ID tracking performance, with an accuracy of 44.1%, an IDF1 of 11.0%, an MOTP of 0.393, and an IDSW of 1. In contrast, the FairMOT+YOLOv5s model recorded an accuracy of 43.8%, an IDF1 of 14.6%, an MOTP of 0.400, and an IDSW of 10. The results of this study indicate that the FairMOT+YOLOv5s model demonstrated higher IDF1 and MOTP scores compared to the FairMOT+DLA-34 model, while the FairMOT+DLA-34 model showed superior performance in ID tracking accuracy and had fewer ID switches.

In this study, we identified the different causes of odor problems and their associated discomfort. We also recognized the significance of public health and environmental concerns. To address odor issues, it is vital to conduct precise analysis and comprehend the root causes. We suggested a hybrid model of a Convolutional Neural Network (CNN) and Transformer called the CNN–Transformer to tackle this challenge and assessed its effectiveness. We utilized a dataset containing 120,000 samples of odor to compare the performance of CNN+LSTM, CNN, LSTM, and ELM models. The experimental results show that the CNN+LSTM hybrid model has an accuracy of 89.00%, precision of 89.41%, recall of 91.04%, F1-score of 90.22%, and RMSE of 0.28, with a large prediction error. The CNN+Transformer hybrid model had an accuracy of 96.21%, precision and recall of 94.53% and 94.16%, F1-score of 94.35%, and RMSE of 0.27, showing a low prediction error. The CNN model had an accuracy of 87.19%, precision and recall of 89.41% and 91.04%, F1-score of 90.22%, and RMSE of 0.23, showing a low prediction error. The LSTM model had an accuracy of 95.00%, precision and recall of 92.55% and 94.17%, F1-score of 92.33%, and RMSE of 0.03, indicating a very low prediction error. The ELM model performed poorly with an accuracy of 85.50%, precision and recall of 85.26% and 85.19%, respectively, and F1-score and RMSE of 85.19% and 0.31, respectively. This study confirms the suitability of the CNN–Transformer hybrid model for odor analysis and highlights its excellent predictive performance. The employment of this model is expected to be advantageous in addressing odor problems and mitigating associated public health and environmental concerns.

Lithium-ion batteries are widely used in electric vehicles, smartphones, and energy storage devices due to their high power and light weight. The goal of this study is to predict the remaining capacity of a lithium-ion battery and evaluate its performance through three machine learning models: linear regression, decision tree, and random forest, and two deep learning models: neural network and ensemble model. Mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R-squared), and root mean squared error (RMSE) were used to measure prediction accuracy. For the evaluation of the artificial intelligence model, the dataset was downloaded and integrated with measurement data of the CS2 lithium-ion battery provided by the University of Maryland College of Engineering. As a result of the study, the RMSE of the linear regression model was 0.045, the decision tree model was 0.038, the random forest model was 0.034, the neural network model was 0.032, and the ensemble model was 0.030. According to the measured values, the ensemble model showed the best predictive performance, followed by the neural network model. Decision tree and random forest models also showed very good performance, and the linear regression model showed relatively poor predictive performance compared to the other models.

X-ray inspection equipment is divided into small baggage inspection equipment and large cargo inspection equipment. In the case of inspection using X-ray scanning equipment, it is possible to identify the contents of goods, unauthorized transport, or hidden goods in real-time by-passing cargo through X-rays without opening it. In this paper, we propose a system for detecting dangerous objects in X-ray images using the Cascade Region-based Convolutional Neural Network (Cascade R-CNN) model, and the data used for learning consists of dangerous goods, storage media, firearms, and knives. In addition, to minimize the overfitting problem caused by the lack of data to be used for artificial intelligence (AI) training, data samples are increased by using the CP (copy-paste) algorithm on the existing data. It also solves the data labeling problem by mixing supervised and semi-supervised learning. The four comparative models to be used in this study are Faster Region-based Convolutional Neural Networks Residual2 Network-101 (Faster R-CNN_Res2Net-101) supervised learning, Cascade R-CNN_Res2Net-101_supervised learning, Cascade Region-based Convolutional Neural Networks Composite Backbone Network V2 (CBNetV2) Network-101 (Cascade R-CNN_ CBNetV2Net-101)_supervised learning, and Cascade R-CNN_ CBNetV2-101_semi-supervised learning which are then compared and evaluated. As a result of comparing the performance of the four models in this paper, in case of Cascade R-CNN_CBNetV2-101_ semi-supervised learning, Average Precision (AP) (Intersection over Union (IoU) = 0.5): 0.7%, AP (IoU = 0.75): 1.0% than supervised learning, Recall: 0.8% higher.

This study aims to detect and prevent greening disease in citrus trees using a deep neural network. The process of collecting data on citrus greening disease is very difficult because the vector pests are too small. In this paper, since the amount of data collected for deep learning is insufficient, we intend to use the efficient feature extraction function of the neural network based on the Transformer algorithm. We want to use the Cascade Region-based Convolutional Neural Networks (Cascade R-CNN) Swin model, which is a mixture of the transformer model and Cascade R-CNN model to detect greening disease occurring in citrus. In this paper, we try to improve model safety by establishing a linear relationship between samples using Mixup and Cutmix algorithms, which are image processing-based data augmentation techniques. In addition, by using the ImageNet dataset, transfer learning, and stochastic weight averaging (SWA) methods, more accuracy can be obtained. This study compared the Faster Region-based Convolutional Neural Networks Residual Network101 (Faster R-CNN ResNet101) model, Cascade Region-based Convolutional Neural Networks Residual Network101 (Cascade R-CNN-ResNet101) model, and Cascade R-CNN Swin Model. As a result, the Faster R-CNN ResNet101 model came out as Average Precision (AP) (Intersection over Union (IoU)=0.5): 88.2%, AP(IoU = 0.75): 62.8%, Recall: 68.2%, and the Cascade R-CNN ResNet101 model was AP(IoU = 0.5): 91.5%, AP (IoU = 0.75): 67.2%, Recall: 73.1%. Alternatively, the Cascade R-CNN Swin Model showed AP (IoU = 0.5): 94.9%, AP (IoU = 0.75): 79.8% and Recall: 76.5%. Thus, the Cascade R-CNN Swin Model showed the best results for detecting citrus greening disease.

This paper presents an interval-based LDA (Linear Discriminant Analysis) algorithm for individual verification using ECG (Electrocardiogram). In this algorithm, at first, unwanted noise and power-line interference are removed from the ECG signal. Then, the autocorrelation profile (ACP) of the ECG signal, which is a mathematical representation of the degree of similarity between a given time series and a lagged version of itself over successive time intervals, is calculated. Finally, the interval-based LDA algorithm is applied to extract unique individual feature vectors that represent distance and angle characteristics on short ACP segments. These feature vectors are used during the processes of enrollment and verification of individual identification. To validate our algorithm, we conducted experiments using the MIT-BIH ECG and achieved EERs (Equal Error Rate) of 0.143%, showing that the proposed algorithm is practically effective and robust in verifying the individual’s identity.