• Smart Media Journal
• /
• v.12 no.11
• /
• pp.77-85
• /
• 2023

After COVID-19, communication through online platforms has increased, leading to an accumulation of massive amounts of conversational text data. With the growing importance of summarizing this text data to extract meaningful information, there has been active research on deep learning-based abstractive summarization. However, conversational data, compared to structured texts like news articles, often contains missing or transformed information, necessitating consideration from multiple perspectives due to its unique characteristics. In particular, vocabulary omissions and unrelated expressions in the conversation can hinder effective summarization. Therefore, in this study, we restructured by considering the characteristics of Korean conversational data, fine-tuning a pre-trained text summarization model based on KoBART, and improved conversation data summary perfomance through a refining operation to remove redundant elements from the summary. By restructuring the sentences based on the order of utterances and extracting a central speaker, we combined methods to restructure the conversation around them. As a result, there was about a 4 point improvement in the Rouge-1 score. This study has demonstrated the significance of our conversation restructuring approach, which considers the characteristics of dialogue, in enhancing Korean conversation summarization performance.

• The Journal of the Convergence on Culture Technology
• /
• v.10 no.4
• /
• pp.639-644
• /
• 2024

Recent developments in the use of drones on battlefields, extending beyond reconnaissance to firepower support, have greatly increased the importance of technologies for early automatic drone identification. In this study, to identify an effective image classification model that can distinguish drones from other aerial targets of similar size and appearance, such as birds and balloons, we utilized a dataset of 3,600 images collected from the internet. We adopted a transfer learning approach that combines the feature extraction capabilities of three pre-trained convolutional neural network models (VGG16, ResNet50, InceptionV3) with an additional classifier. Specifically, we conducted a comparative analysis of the performance of these three pre-trained models to determine the most effective one. The results showed that the InceptionV3 model achieved the highest accuracy at 99.66%. This research represents a new endeavor in utilizing existing convolutional neural network models and transfer learning for drone identification, which is expected to make a significant contribution to the advancement of drone identification technologies.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.6
• /
• pp.1818-1832
• /
• 2022

Violence in the Internet era poses a new challenge to the current counter-riot work, and according to research and analysis, most of the violent incidents occurring are related to the dissemination of violence images. The use of the popular deep learning neural network to automatically analyze the massive amount of images on the Internet has become one of the important tools in the current counter-violence work. This paper focuses on the use of transfer learning techniques and the introduction of an attention mechanism to the residual network (ResNet) model for the classification and identification of violence images. Firstly, the feature elements of the violence images are identified and a targeted dataset is constructed; secondly, due to the small number of positive samples of violence images, pre-training and attention mechanisms are introduced to suggest improvements to the traditional residual network; finally, the improved model is trained and tested on the constructed dedicated dataset. The research results show that the improved network model can quickly and accurately identify violence images with an average accuracy rate of 92.20%, thus effectively reducing the cost of manual identification and providing decision support for combating rebel organization activities.

• Annual Conference of KIPS
• /
• 2023.05a
• /
• pp.687-688
• /
• 2023

Bone age assessment is a crucial task in pediatric radiology for assessing growth and development in children. In this paper, we explore the potential of Vision Transformer, a state-of-the-art deep learning model, for bone age assessment using X-ray images. We generate heatmap outputs using a pre-trained Vision Transformer model on a publicly available dataset of hand X-ray images and show that the model tends to focus on the overall hand and only the bone part of the image, indicating its potential for accurately identifying the regions of interest for bone age assessment without the need for pre-processing to remove background noise. We also suggest two methods for extracting the region of interest from the heatmap output. Our study suggests that Vision Transformer holds great potential for bone age assessment using X-ray images, as it can provide accurate and interpretable output that may assist radiologists in identifying potential abnormalities or areas of interest in the X-ray image.

• Annual Conference of KIPS
• /
• 2020.11a
• /
• pp.1014-1017
• /
• 2020

Many natural language processing (NLP) models utilize pre-trained word embeddings to leverage latent information. One of the most successful word embedding model is the Skip-gram (SG). In this paper, we propose a Skipgram drop (SG-Drop) model, which is a variation of the SG model. The SG-Drop model is designed to reduce training time efficiently. Furthermore, the SG-Drop allows controlling training time with its hyperparameter. It could train word embedding faster than reducing training epochs while better preserving the quality.

• Journal of Convergence for Information Technology
• /
• v.11 no.9
• /
• pp.28-34
• /
• 2021

In the case of object detection using deep learning, both accuracy and real-time are required. However, it is difficult to use a deep learning model that processes a large amount of data in a limited resource environment. To solve this problem, this paper proposes an object detection model for small embedded devices. Unlike the general detection model, the model size was minimized by using a structure in which the pre-trained feature extractor was removed. The structure of the model was designed by repeatedly stacking lightweight convolution blocks. In addition, the number of region proposals is greatly reduced to reduce detection overhead. The proposed model was trained and evaluated using the public dataset PASCAL VOC. For quantitative evaluation of the model, detection performance was measured with average precision used in the detection field. And the detection speed was measured in a Raspberry Pi similar to an actual embedded device. Through the experiment, we achieved improved accuracy and faster reasoning speed compared to the existing detection method.

• Journal of Biomedical Engineering Research
• /
• v.43 no.6
• /
• pp.434-440
• /
• 2022

Liver transplantation is a critical used treatment method for patients with end-stage liver disease. The number of cases of living donor liver transplantation is increasing due to the imbalance in needs and supplies for brain-dead organ donation. As a result, the importance of the accuracy of the donor's suitability evaluation is also increasing rapidly. To measure the donor's liver volume accurately is the most important, that is absolutely necessary for the recipient's postoperative progress and the donor's safety. Therefore, we propose liver segmentation in abdominal CT images from pre-operation, POD 7, and POD 63 with a two-dimensional U-Net. In addition, we introduce an algorithm to measure the volume of the segmented liver and measure the hepatectomy rate and regeneration rate of pre-operation, POD 7, and POD 63. The performance for the learning model shows the best results in the images from pre-operation. Each dataset from pre-operation, POD 7, and POD 63 has the DSC of 94.55 ± 9.24%, 88.40 ± 18.01%, and 90.64 ± 14.35%. The mean of the measured liver volumes by trained model are 1423.44 ± 270.17 ml in pre-operation, 842.99 ± 190.95 ml in POD 7, and 1048.32 ± 201.02 ml in POD 63. The donor's hepatectomy rate is an average of 39.68 ± 13.06%, and the regeneration rate in POD 63 is an average of 14.78 ± 14.07%.

• Journal of Platform Technology
• /
• v.11 no.4
• /
• pp.3-12
• /
• 2023

In the post-COVID era, the importance of quarantine measures is greatly emphasized, and accordingly, research related to the detection of mask wearing conditions and prevention of other infectious diseases using deep learning is being conducted. However, research on the detection and tracking of visitors to cultural facilities to prevent the spread of diseases is equally important, so research on this should be conducted. In this paper, a convolutional neural network-based object detection model is trained through transfer learning using a pre-collected dataset. The weights of the trained detection model are then applied to a multi-object tracking model to monitor visitors. The visitor detection model demonstrates results with a precision of 96.3%, recall of 85.2%, and an F1-score of 90.4%. Quantitative results of the tracking model include a MOTA (Multiple Object Tracking Accuracy) of 65.6%, IDF1 (ID F1 Score) of 68.3%, and HOTA (Higher Order Tracking Accuracy) of 57.2%. Furthermore, a qualitative comparison with other multi-object tracking models showcased superior results for the model proposed in this paper. The research of this paper can be applied to the hygiene systems within cultural facilities in the post-COVID era.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.10
• /
• pp.9-16
• /
• 2023

Technological advancement is being advanced in sports such as electronic protection of taekwondo competition and VAR of soccer. However, a person judges and guides the posture by looking at the posture, so sometimes a judgment dispute occurs at the site of the competition in Taekwondo Poomsae. This study proposes an artificial intelligence model that can more accurately judge and evaluate Taekwondo movements using artificial intelligence. In this study, after pre-processing the photographed and collected data, it is separated into train, test, and validation sets. The separated data is trained by applying each model and conditions, and then compared to present the best-performing model. The models under each condition compared the values of loss, accuracy, learning time, and top-n error, and as a result, the performance of the model trained under the conditions using ResNet50 and Adam was found to be the best. It is expected that the model presented in this study can be utilized in various fields such as education sites and competitions.

• Journal of IKEEE
• /
• v.24 no.4
• /
• pp.1011-1016
• /
• 2020

We propose a framework called Stacked Gated Recurrent Unit - Infrequent Residual Analysis (SG-IRA) that detects anomalies in time-series data that can be trained on streams of raw sensor data without any pre-labeled dataset. To enable such unsupervised learning, SG-IRA includes an estimation model that uses a stacked Gated Recurrent Unit (GRU) structure and an analysis method that detects anomalies based on the difference between the estimated value and the actual measurement (residual). SG-IRA's residual analysis method dynamically adapts the detection threshold from the population using frequency analysis, unlike the baseline model that relies on a constant threshold. In this paper, SG-IRA is evaluated using the industrial control systems (ICS) datasets. SG-IRA improves the detection performance (F1 score) by 5.9% compared to the baseline model.