• Korean Journal of Remote Sensing
• /
• 제38권6_2호
• /
• pp.1663-1676
• /
• 2022

Recently, a number of deep-learning based land cover segmentation studies have been introduced. Some studies denoted that the performance of land cover segmentation deteriorated due to insufficient training data. In this study, we verified the improvement of land cover segmentation performance through data augmentation. U-Net was implemented for the segmentation model. And 2020 satellite-derived landcover dataset was utilized for the study data. The pixel accuracies were 0.905 and 0.923 for U-Net trained by original and augmented data respectively. And the mean F1 scores of those models were 0.720 and 0.775 respectively, indicating the better performance of data augmentation. In addition, F1 scores for building, road, paddy field, upland field, forest, and unclassified area class were 0.770, 0.568, 0.433, 0.455, 0.964, and 0.830 for the U-Net trained by original data. It is verified that data augmentation is effective in that the F1 scores of every class were improved to 0.838, 0.660, 0.791, 0.530, 0.969, and 0.860 respectively. Although, we applied data augmentation without considering class balances, we find that data augmentation can mitigate biased segmentation performance caused by data imbalance problems from the comparisons between the performances of two models. It is expected that this study would help to prove the importance and effectiveness of data augmentation in various image processing fields.

• Korean Journal of Remote Sensing
• /
• 제38권6_4호
• /
• pp.1925-1934
• /
• 2022

Monitoring nuclear activity for inaccessible areas using remote sensing technology is essential for nuclear non-proliferation. In recent years, deep learning has been actively used to detect nuclear-activity-related small objects. However, high-resolution satellite imagery containing small objects can result in class imbalance. As a result, there is a performance degradation problem in detecting small objects. Therefore, this study aims to improve detection accuracy by analyzing the effect of the ratio of small objects related to nuclear activity in the input data for the performance of the deep learning model. To this end, six case datasets with different ratios of small object pixels were generated and a U-Net model was trained for each case. Following that, each trained model was evaluated quantitatively and qualitatively using a test dataset containing various types of small object classes. The results of this study confirm that when the ratio of object pixels in the input image is adjusted, small objects related to nuclear activity can be detected efficiently. This study suggests that the performance of deep learning can be improved by adjusting the object pixel ratio of input data in the training dataset.

• Journal of the Korean Society of Radiology
• /
• 제16권6호
• /
• pp.697-702
• /
• 2022

For imaging in positron emission tomography(PET), it is necessary to measure the position of the scintillation pixel interacting with the gamma rays incident on the detector. To this end, in the conventional system, a flood image of the scintillation pixel is obtained, the imaged area of each scintillation pixel is separated, and the position of the scintillation pixel is specified and acquired as a digital signal. In this study, a deep learning method was applied based on the signal formed by the photosensor of the detector, and a method was developed to directly acquire a digital signal without going through various procedures. DETECT2000 simulation was performed to verify this and evaluate the accuracy of position measurement. A detector was constructed using a 6 × 6 scintillation pixel array and a 4 × 4 photosensor, and a gamma ray event was generated at the center of the scintillation pixel and summed into four channels of signals through the Anger equation. After training the deep learning model using the acquired signal, the positions of gamma-ray events that occurred in different depth directions of the scintillation pixel were measured. The results showed accurate results at every scintillation pixel and position. When the method developed in this study is applied to the PET detector, it will be possible to measure the position of the scintillation pixel with a digital signal more conveniently.

• Journal of Intelligence and Information Systems
• /
• 제28권3호
• /
• pp.119-137
• /
• 2022

Recent recommendation system studies apply various deep learning models to represent user and item interactions better. One of the noteworthy studies is ONCF(Outer product-based Neural Collaborative Filtering) which builds a two-dimensional interaction map via outer product and employs CNN (Convolutional Neural Networks) to learn high-order correlations from the map. However, ONCF has limitations in recommendation performance due to the problems with CNN and the absence of side information. ONCF using CNN has an inductive bias problem that causes poor performances for data with a distribution that does not appear in the training data. This paper proposes to employ a Vision Transformer (ViT) instead of the vanilla CNN used in ONCF. The reason is that ViT showed better results than state-of-the-art CNN in many image classification cases. In addition, we propose a new architecture to reflect side information that ONCF did not consider. Unlike previous studies that reflect side information in a neural network using simple input combination methods, this study uses an independent auxiliary classifier to reflect side information more effectively in the recommender system. ONCF used a single latent vector for user and item, but in this study, a channel is constructed using multiple vectors to enable the model to learn more diverse expressions and to obtain an ensemble effect. The experiments showed our deep learning model improved performance in recommendation compared to ONCF.

• KIPS Transactions on Software and Data Engineering
• /
• 제11권8호
• /
• pp.325-330
• /
• 2022

This paper conducts the collection and experiment of datasets for deep learning research on sign language such as sign language recognition, sign language translation, and sign language segmentation for Korean sign language. There exist difficulties for deep learning research of sign language. First, it is difficult to recognize sign languages since they contain multiple modalities including hand movements, hand directions, and facial expressions. Second, it is the absence of training data to conduct deep learning research. Currently, KETI dataset is the only known dataset for Korean sign language for deep learning. Sign language datasets for deep learning research are classified into two categories: Isolated sign language and Continuous sign language. Although several foreign sign language datasets have been collected over time. they are also insufficient for deep learning research of sign language. Therefore, we attempted to collect a large-scale Korean sign language dataset and evaluate it using a baseline model named TSPNet which has the performance of SOTA in the field of sign language translation. The collected dataset consists of a total of 11,402 image and text. Our experimental result with the baseline model using the dataset shows BLEU-4 score 3.63, which would be used as a basic performance of a baseline model for Korean sign language dataset. We hope that our experience of collecting Korean sign language dataset helps facilitate further research directions on Korean sign language.

• The Korean Journal of Applied Statistics
• /
• 제36권2호
• /
• pp.101-114
• /
• 2023

The data augmentation technique is effectively used to solve the problem of overfitting the model by allowing the training dataset to be viewed from various perspectives. In addition to image augmentation techniques such as rotation, cropping, horizontal flip, and vertical flip, occlusion-based data augmentation methods such as Cutmix and Cutout have been proposed. For models based on speech data, it is possible to use an occlusion-based data-based augmentation technique after converting a 1D speech signal into a 2D spectrogram. In particular, SpecAugment is an occlusion-based augmentation technique for speech spectrograms. In this study, we intend to compare and study data augmentation techniques that can be used in the problem of false-voice detection. Using data from the ASVspoof2017 and ASVspoof2019 competitions held to detect fake audio, a dataset applied with Cutout, Cutmix, and SpecAugment, an occlusion-based data augmentation method, was trained through an LCNN model. All three augmentation techniques, Cutout, Cutmix, and SpecAugment, generally improved the performance of the model. In ASVspoof2017, Cutmix, in ASVspoof2019 LA, Mixup, and in ASVspoof2019 PA, SpecAugment showed the best performance. In addition, increasing the number of masks for SpecAugment helps to improve performance. In conclusion, it is understood that the appropriate augmentation technique differs depending on the situation and data.

• Journal of IKEEE
• /
• 제27권3호
• /
• pp.345-349
• /
• 2023

In this paper, we propose a study on the development of deep learning structure for defective pixel detection of next-generation smart LED display board using imaging device. In this research, a technique utilizing imaging devices and deep learning is introduced to automatically detect defects in outdoor LED billboards. Through this approach, the effective management of LED billboards and the resolution of various errors and issues are aimed. The research process consists of three stages. Firstly, the planarized image data of the billboard is processed through calibration to completely remove the background and undergo necessary preprocessing to generate a training dataset. Secondly, the generated dataset is employed to train an object recognition network. This network is composed of a Backbone and a Head. The Backbone employs CSP-Darknet to extract feature maps, while the Head utilizes extracted feature maps as the basis for object detection. Throughout this process, the network is adjusted to align the Confidence score and Intersection over Union (IoU) error, sustaining continuous learning. In the third stage, the created model is employed to automatically detect defective pixels on actual outdoor LED billboards. The proposed method, applied in this paper, yielded results from accredited measurement experiments that achieved 100% detection of defective pixels on real LED billboards. This confirms the improved efficiency in managing and maintaining LED billboards. Such research findings are anticipated to bring about a revolutionary advancement in the management of LED billboards.

• The Journal of Korean Philosophical History
• /
• 제27호
• /
• pp.417-446
• /
• 2009

This study was to review the features of the preventive dimension in connection with adolescent deviation on the basis of the morality and ethics held by Confucian doctrine. To find solutions to the problems of adolescent deviation is never easy. As adolescent deviation always does occur, it is important to consider the methods that can minimize and prevent it. The traditional society of Korea laid weight on the education and training in the aspect of preventive measure against such adolescent deviation by emphasizing moral edification and realization of spiritual understanding for it. In this paper, the researcher tried to understand the problem situations by examining the image of such deviation and its type as well as the method on response thereto targeting the young generation of Korea. In addition, the researcher analyzed how the adolescent was recognized in the traditional society that was established on the Confucian values, and moral standards that applied to them, and the process of education as well. Through the moral concepts of Confucianism that were revealed in the Doctrine of the Mean (中庸, pronounced 'Jungyong' in Korean) in particular, the researcher sought the possibility of education on morality and ethics that will be able to prevent adolescent deviation. This study suggests that the morality and ethics held by Confucian doctrine can prevent adolescent deviation and open a new horizon of ethics education.

• Journal of Korea Water Resources Association
• /
• 제55권spc1호
• /
• pp.1177-1185
• /
• 2022

The water and sewage system is an infrastructure that provides safe and clean water to people. In particular, since the water and sewage pipelines are buried underground, it is very difficult to detect system defects. For this reason, the diagnosis of pipelines is limited to post-defect detection, such as system diagnosis based on the images taken after taking pictures and videos with cameras and drones inside the pipelines. Therefore, real-time detection technology of pipelines is required. Recently, pipeline diagnosis technology using advanced equipment and artificial intelligence techniques is being developed, but AI-based defect detection technology requires a variety of learning data because the types and numbers of defect data affect the detection performance. Therefore, in this study, various defect scenarios are implemented using 3D printing model to improve the detection performance when detecting defects in pipelines. Afterwards, the collected images are performed to pre-processing such as classification according to the degree of risk and labeling of objects, and real-time defect detection is performed. The proposed technique can provide real-time feedback in the pipeline defect detection process, and it would be minimizing the possibility of missing diagnoses and improve the existing water and sewerage pipe diagnosis processing capability.

• The Journal of the Acoustical Society of Korea
• /
• 제42권4호
• /
• pp.329-344
• /
• 2023

In this study, we comprehensively analyze the generalization performance of various deep learning-based active sonar target classifiers when applied to small and imbalanced active sonar datasets. To generate the active sonar datasets, we use data from two different oceanic experiments conducted at different times and ocean. Each sample in the active sonar datasets is a time-frequency domain image, which is extracted from audio signal of contact after the detection process. For the comprehensive analysis, we utilize 22 Convolutional Neural Networks (CNN) models. Two datasets are used as train/validation datasets and test datasets, alternatively. To calculate the variance in the output of the target classifiers, the train/validation/test datasets are repeated 10 times. Hyperparameters for training are optimized using Bayesian optimization. The results demonstrate that shallow CNN models show superior robustness and generalization performance compared to most of deep CNN models. The results from this paper can serve as a valuable reference for future research directions in deep learning-based active sonar target classification.