KJCKorea
Journal Central

In this paper, we propose an infrared camera anomaly detection system based on deep learning. In general industrial sites, infrared cameras are used for facility management. However, if something goes wrong with the infrared camera, recording stops and the status of the facility is unknown. Therefore, monitoring is necessary to prepare for this. Currently, only manpower is monitoring the camera. However, there are limitations such as human error and mass inspection. To solve this limitation, we compare performance using AlexNet, VGGNet, ResNet, which are types of deep learning, and propose an automated monitoring method with better performance by comparing the captured original image and the histogram extracted image as a preprocessing process. The data set was collected and used by KEPCO's infrared night vision camera, and the results of training the data set with the VGGNet model and predicting the results were the best. The result is that the input image is judged to be normal/abnormal for the image captured at a certain period, and the model performance came out as Accuracy 0.97, Precision 0.97, Recall 0.97, F1-score 0.97. In addition, it was confirmed that the data set preprocessed with the histogram performed better.

This study developed and analyzed ensemble learning-based prediction models for forecasting attendance in the KBO League. Using KBO League data from 2022 to 2024, we collected variables such as team rankings, winning rates, consecutive wins/losses, search volume, stadiums, and home/away games, with the attendance ratio compared to stadium capacity set as the target variable. In the data preprocessing phase, Monday games were excluded, and the home/away attendance ratio was set to 7:3 to enhance model realism. Among various ensemble models compared, including Linear Regression, Random Forest, XGBoost, and LightGBM, the LightGBM model showed the best performance with an RMSE of 8.39 and R² Score of 0.783. Feature importance analysis revealed that online search volume (28.17%) and winning rate (25.17%) had the most significant impact on attendance, while team (10.57%) and day of the week (9.73%) also showed meaningful influence. Additionally, SHAP (SHapley Additive exPlanations) analysis provided insights into the directional impact of each variable on predictions, particularly revealing that the home/away factor had a stronger influence than expected through interactions with other variables. This study is significant in providing a practical prediction model that can assist KBO teams in establishing attendance strategies and making marketing decisions.

As the K-POP industry continues to grow in the global market, the importance of choreography originality and copyright protection is increasing. However, existing methods of recording choreography primarily rely on videos and documents, which have limitations in systematic preservation and protection. To address this issue, this study aims to develop an AI-powered system that automatically analyzes and converts K-POP choreography into a dance notation format. In this study, video frames are extracted as images, and Pose is utilized to capture poses from each frame. These poses are then sequentially arranged, similar to musical notation, to systematically represent the choreography. Additionally, emotion icons are incorporated into the notation to visually convey the emotional expressions of the movements, making interpretation more intuitive. Through this research, we aim to develop an AI-based automated dance notation system, providing a structured approach to recording and preserving K-POP choreography while contributing to the entertainment industry.