• Journal of Internet Computing and Services
• /
• v.21 no.6
• /
• pp.23-31
• /
• 2020

This study was carried out to generate various images of railroad surfaces with random defects as training data to be better at the detection of defects. Defects on the surface of railroads are caused by various factors such as friction between track binding devices and adjacent tracks and can cause accidents such as broken rails, so railroad maintenance for defects is necessary. Therefore, various researches on defect detection and inspection using image processing or machine learning on railway surface images have been conducted to automate railroad inspection and to reduce railroad maintenance costs. In general, the performance of the image processing analysis method and machine learning technology is affected by the quantity and quality of data. For this reason, some researches require specific devices or vehicles to acquire images of the track surface at regular intervals to obtain a database of various railway surface images. On the contrary, in this study, in order to reduce and improve the operating cost of image acquisition, we constructed the 'Defective Railroad Surface Regeneration Model' by applying the methods presented in the related studies of the Generative Adversarial Network (GAN). Thus, we aimed to detect defects on railroad surface even without a dedicated database. This constructed model is designed to learn to generate the railroad surface combining the different railroad surface textures and the original surface, considering the ground truth of the railroad defects. The generated images of the railroad surface were used as training data in defect detection network, which is based on Fully Convolutional Network (FCN). To validate its performance, we clustered and divided the railroad data into three subsets, one subset as original railroad texture images and the remaining two subsets as another railroad surface texture images. In the first experiment, we used only original texture images for training sets in the defect detection model. And in the second experiment, we trained the generated images that were generated by combining the original images with a few railroad textures of the other images. Each defect detection model was evaluated in terms of 'intersection of union(IoU)' and F1-score measures with ground truths. As a result, the scores increased by about 10~15% when the generated images were used, compared to the case that only the original images were used. This proves that it is possible to detect defects by using the existing data and a few different texture images, even for the railroad surface images in which dedicated training database is not constructed.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.05a
• /
• pp.665-666
• /
• 2022

반려동물의 이상행동 탐지를 위한 센서 데이터를 수집하는 과정에서 발생하는 시간과 비용의 문제로 인해 데이터 증강이 요구되고 있다. 본 논문에서는 통계적 변형과 GAN 기반의 데이터 증강을 통해 반려동물의 정상행동과 이상행동으로 분류하는 방법을 제안한다. 통계적 변형은 회전, 순열, 조합 등을 이용하며, GAN을 통해 원본 데이터에 노이즈가 포함된 유사한 데이터를 생성한다. 증강된 모든 데이터는 원본 데이터와 함께 학습 데이터로 사용한다. 최종적으로, LSTM의 단점을 보완한 Convolutional LSTM 모델을 통해 반려동물의 정상행동 인식의 범주를 넓혀 보다 정확한 이상행동을 인식하고자 한다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.11a
• /
• pp.4-7
• /
• 2022

기존 합성곱 신경망 기반의 잡음 제거 네트워크들은 학습을 위한 noisy-clean 데이터 쌍을 필요로 한다. 하지만 실제 카메라 잡음의 경우, 잡음에 대한 깨끗한 원본 영상을 얻는 것은 불가능하거나 많은 비용이 소모된다. 따라서 이러한 방법을 해결하기 위하여 원본 영상 없이 잡음 영상만으로만 잡음 제거 네트워크를 학습하는 방법들이 제안되어왔다. 그 중 카메라 잡음 영상을 처리하기 위한 대표적인 방법으로 학습과 추론에서 비대칭적인 downsampling을 사용하는 AP-BSN이 제안되었다. 본 논문에서는 Functional neural network를 AP-BSN 알고리즘에 적용하여 다양한 downsampling ratio에 대응되는 하나의 네트워크를 학습하였다. 이를 통해 기존 hyperparameter로 사용되던 downsampling ratio에 대한 결과를 하나의 네트워크에서 분석 및 확인하였다. 또한 해당 파라미터를 조절함으로써 다양한 잡음 제거 후보들을 추출하고 사용자가 원하는 잡음 제거 정도를 조정할 수 있도록 하였다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2020.11a
• /
• pp.838-841
• /
• 2020

본 논문은 영상 속 폭력행위를 인식하기 위해 3 차원 컨벌루션을 활용하여 원본 영상과 스켈레톤(skeleton)영상으로부터 추출한 시각 및 움직임 정보를 동시에 활용하는 2-스트림 구조의 폭력상황 인식기를 제안한다. 제안된 폭력상황 인식기에서는 수평, 수직 방향의 큰 움직임이 많이 나타나는 폭력영상의 특성을 활용하기위해 각 방향의 특성을 독립적으로 학습할 수 있는 split-FAST 3차원 컨벌루션을 활용하고, 3 차원 Attention 을 적용하여 시각 및 움직임 정보 추출 시 영상의 중요지역을 중점적으로 반영하도록 함으로써 촬영 기기의 이동 또는 여러 사람의 뒤엉킴 등으로 영상의 시점 변화나 상황 변화가 잦은 경우에도 강인한 성능을 가질 수 있도록 하였다. 또한 기존의 연구들과 달리 비제약적인 환경에서 CCTV, 모바일 카메라 등으로 촬영된 실제 영상들로 구성된 RLVS 데이터셋을 학습 데이터로 사용함으로써 실제의 폭력 행위를 잘 인식할 수 있도록 하였다. RLVS 를 이용한 평가 실험에서 제안된 폭력상황 인식기가 약 92%의 인식 정확도를 얻었다.

• Journal of Korea Society of Industrial Information Systems
• /
• v.27 no.4
• /
• pp.19-27
• /
• 2022

In general, the performance of ML(Machine Learning) application is determined by various factors such as the type of ML model, the size of model (number of parameters), hyperparameters setting during the training, and training data. In particular, the recognition accuracy of ML may be deteriorated or experienced overfitting problem if the amount of dada used for training is insufficient. Existing studies focusing on image recognition have widely used open datasets for training and evaluating the proposed ML models. However, for specific applications where the sensor used, the target of recognition, and the recognition situation are different, it is necessary to build the dataset manually. In this case, the performance of ML largely depends on the quantity and quality of the data. In this paper, training data used for motion recognition application is augmented using the kernel density estimation algorithm which is a type of non-parametric estimation method. We then compare and analyze the recognition accuracy of a ML application by varying the number of original data, kernel types and augmentation rate used for data augmentation. Finally experimental results show that the recognition accuracy is improved by up to 14.31% when using the narrow bandwidth Tophat kernel.

• Journal of the Korea Society for Simulation
• /
• v.33 no.2
• /
• pp.27-35
• /
• 2024

We aimed to shed light on the issue of perpetuating societal disparities by analyzing the impact of inherent biases present in datasets used for training artificial intelligence models on the predictions generated by Artificial Intelligence(AI). Therefore, to examine the influence of data bias on AI models, we constructed an original dataset containing biases related to gender wage gaps and subsequently created a de-identified dataset. Additionally, by utilizing the decision tree algorithm, we compared the outputs of AI models trained on both the original and de-identified datasets, aiming to analyze how data de-identification affects the biases in the results produced by artificial intelligence models. Through this, our goal was to highlight the significant role of data de-identification not only in safeguarding individual privacy but also in addressing biases within the data.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.234-236
• /
• 2023

Edge AI 환경에서의 DNNs 학습 방법 중 하나인 분할 학습은 모델 전도 공격으로 인해 입력 데이터의 프라이버시가 노출될 수 있다. 본 논문에서는 분할 학습 환경에서의 모델 전도 공격에 대한 기존 방어 기술들의 한계점을 회피할 수 있는 TrapMI 기술을 제안하고, 이를 통해 입력 이미지를 원 본 데이터 세트의 도메인에서 특정 타겟 이미지 도메인으로 이동시킴으로써 이미지 복원의 가능성을 최소화시킨다. 추가적으로, 테스트 과정에서 타겟 이미지의 정보를 알 수 없는 제약을 회피하기 위해 AutoGenerator를 구축한 후 실험을 통해 원본 데이터 보호 성능을 검증한다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.203-208
• /
• 2023

본 논문에서는 대규모 언어 모델의 인컨텍스트 러닝과 프롬프팅을 활용하여 수치 추론 태스크 데이터셋을 효과적으로 증강시킬 수 있는 방법론을 제안한다. 또한 모델로 하여금 수치 추론 데이터의 이해를 도울 수 있는 전처리와 요구사항을 만족하지 못하는 결과물을 필터링 하는 검증 단계를 추가하여 생성되는 데이터의 퀄리티를 보장하고자 하였다. 이렇게 얻어진 증강 절차를 거쳐 증강을 진행한 뒤 추론용 모델 학습을 통해 다른 증강 방법론보다 우리의 방법론으로 증강된 데이터셋으로 학습된 모델이 더 높은 성능을 낼 수 있음을 보였다. 실험 결과 우리의 증강 데이터로 학습된 모델은 원본 데이터로 학습된 모델보다 모든 지표에서 2%p 이상의 성능 향상을 보였으며 다양한 케이스를 통해 우리의 모델이 수치 추론 학습 데이터의 다양성을 크게 향상시킬 수 있음을 확인하였다.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.32 no.2
• /
• pp.405-416
• /
• 2022

Current AI technology is improving the quality of life by using machine learning based on data. When using machine learning, transmitting distributed data and collecting it in one place goes through a de-identification process because there is a risk of privacy infringement. De-identification data causes information damage and omission, which degrades the performance of the machine learning process and complicates the preprocessing process. Accordingly, Google announced joint learning in 2016, a method of de-identifying data and learning without the process of collecting data into one server. This paper analyzed the effectiveness by comparing the difference between the learning performance of data that went through the de-identification process of K anonymity and differential privacy reproduction data using actual financial data. As a result of the experiment, the accuracy of original data learning was 79% for k=2, 76% for k=5, 52% for k=7, 50% for 𝜖=1, and 82% for 𝜖=0.1, and 86% for Federated learning.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_2
• /
• 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.