• The korean journal of orthodontics
• /
• v.51 no.2
• /
• pp.77-85
• /
• 2021

Objective: To evaluate the accuracy of a multi-stage convolutional neural network (CNN) model-based automated identification system for posteroanterior (PA) cephalometric landmarks. Methods: The multi-stage CNN model was implemented with a personal computer. A total of 430 PA-cephalograms synthesized from cone-beam computed tomography scans (CBCT-PA) were selected as samples. Twenty-three landmarks used for Tweemac analysis were manually identified on all CBCT-PA images by a single examiner. Intra-examiner reproducibility was confirmed by repeating the identification on 85 randomly selected images, which were subsequently set as test data, with a two-week interval before training. For initial learning stage of the multi-stage CNN model, the data from 345 of 430 CBCT-PA images were used, after which the multi-stage CNN model was tested with previous 85 images. The first manual identification on these 85 images was set as a truth ground. The mean radial error (MRE) and successful detection rate (SDR) were calculated to evaluate the errors in manual identification and artificial intelligence (AI) prediction. Results: The AI showed an average MRE of 2.23 ± 2.02 mm with an SDR of 60.88% for errors of 2 mm or lower. However, in a comparison of the repetitive task, the AI predicted landmarks at the same position, while the MRE for the repeated manual identification was 1.31 ± 0.94 mm. Conclusions: Automated identification for CBCT-synthesized PA cephalometric landmarks did not sufficiently achieve the clinically favorable error range of less than 2 mm. However, AI landmark identification on PA cephalograms showed better consistency than manual identification.

• Journal of Institute of Control, Robotics and Systems
• /
• v.13 no.11
• /
• pp.1074-1081
• /
• 2007

This paper presents a method for real-time face detection and tracking which combined Adaboost and Camshift algorithm. Adaboost algorithm is a method which selects an important feature called weak classifier among many possible image features by tuning weight of each feature from learning candidates. Even though excellent performance extracting the object, computing time of the algorithm is very high with window size of multi-scale to search image region. So direct application of the method is not easy for real-time tasks such as multi-task OS, robot, and mobile environment. But CAMshift method is an improvement of Mean-shift algorithm for the video streaming environment and track the interesting object at high speed based on hue value of the target region. The detection efficiency of the method is not good for environment of dynamic illumination. We propose a combined method of Adaboost and CAMshift to improve the computing speed with good face detection performance. The method was proved for real image sequences including single and more faces.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.541-543
• /
• 2022

In recent years, retinal disorders have become a serious health concern. Retinal disorders develop slowly and without obvious signs. To avoid vision deterioration, early detection and treatment are critical. Optical coherence tomography (OCT) is a non-invasive and non-contact medical imaging technique used to acquire informative and high-resolution image of retinal area and underlying layers. Disease signs are difficult to detect because OCT images have many areas which are not related to any disease. In this paper, we present a deep learning-based method to perform multi-label classification on a long-tailed OCT dataset. Our method first extracts the region of interest and then performs the classification task. We achieve 98% accuracy, 92% sensitivity, and 99% specificity on our private OCT dataset. Using the heatmap generated from trained convolutional neural network, our method is more robust and explainable than previous approaches because it focuses on areas that contain disease signs.

• Journal of Institute of Control, Robotics and Systems
• /
• v.5 no.1
• /
• pp.95-104
• /
• 1999

In this paper we propose a fast and precise control algorithm for a mobile robot, which aims at the self-tuning control applying two multi-layered neural networks to the structure of computed torque method. Through this algorithm, the nonlinear terms of external disturbance caused by variable task environments and dynamic model errors are estimated and compensated in real time by a long term neural network which has long learning period to extract the non-linearity globally. A short term neural network which has short teaming period is also used for determining optimal gains of PID compensator in order to come over the high frequency disturbance which is not known a priori, as well as to maintain the stability. To justify the global effectiveness of this algorithm where each of the long term and short term neural networks has its own functions, simulations are peformed. This algorithm can also be utilized to come over the serious shortcoming of neural networks, i.e., inefficiency in real time.

• 제어로봇시스템학회:학술대회논문집
• /
• 1991.10a
• /
• pp.280-285
• /
• 1991

Neural network control has many innovative potentials for fast, accurate and intelligent adaptive control. In this paper, a learning control architecture for the dynamic control of a robot manipulator is developed using inverse dynamic neurocontroller and linear neurocontroher. The inverse dynamic neurocontrouer consists of a MLP (multi-layer perceptron) and the linear neurocontroller consists of SLPs (single layer perceptron). Compared with the previous type of neurocontroller which is using an inverse dynamic neurocontroller and a fixed PD gain controller, proposed architecture shows the superior performance over the previous type of neurocontroller because linear neurocontroller can adapt its gain according to the applied task. This superior performance is tested and verified through the control of PUMA 560. Without any knowledge on the dynamic model, its parameters of a robot , (The robot is treated as a complete black box), the neurocontroller, through practice, gradually and implicitly learns the robot's dynamic properties which is essential for fast and accurate control.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.630-633
• /
• 2022

속성기반 감성 분석은 텍스트 내에 존재하는 속성에 대해 세분화된 감성 분석을 수행하는 과제를 말한다. 세분화된 감성분석을 정확하게 수행하기 위해서는 텍스트에 존재하는 감성 표현과 그것이 수식하는 대상에 대한 정보가 반드시 필요하다. 그리고 순서대로 두 가지 정보는 이후 정보를 텍스트에서 추출하기 위해 중요한 단서가 된다. 따라서 본 논문에서는 KorBERT와 Bi-LSTM을 이용한 단계적 특징을 활용한 다중 작업 학습 모델을 사용하여 한국어 감성 분석 말뭉치의 감성 표현과 대상을 추출하는 작업을 수행하였다. 제안한 모델을 한국어 감성 분석 말뭉치로 학습 및 평가한 결과, 감성 표현 추출 작업의 출력을 추가적인 특성으로 전달하여 대상 추출 작업의 성능을 향상시킬 수 있음을 보였다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2021.06a
• /
• pp.105-108
• /
• 2021

Generative adversarial networks (GANs) have reached a great result at creating the synthesis image, especially in the face generation task. Unlike other deep learning tasks, the input of GANs is usually the random vector sampled by a probability distribution, which leads to unstable training and unpredictable output. One way to solve those problems is to employ the label condition in both the generator and discriminator. CelebA and FFHQ are the two most famous datasets for face image generation. While CelebA contains attribute annotations for more than 200,000 images, FFHQ does not have attribute annotations. Thus, in this work, we introduce a method to learn the attributes from CelebA then predict both soft and hard labels for FFHQ. The evaluated result from our model achieves 0.7611 points of the metric is the area under the receiver operating characteristic curve.

• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.418-420
• /
• 2020

해양에서의 선박사고 발생 횟수는 매년 꾸준히 증가하고 있다. 한국해양안전심판원에서는 이러한 사례들의 판결을 관련 인력들이 공유할 수 있도록 재결서를 제작하여 발간하고 있다. 그러나 선박사고는 2019년 기준 2,971건이 발생하여, 재결서만으로 관련 인력들이 다양한 사건들의 판례를 익히기엔 어려움이 따른다. 따라서 본 논문에서는 문장 표상 기법을 이용한 다중 작업 학습을 이용하여 선박사고의 사고 유형, 적용되는 법령, 형량을 분류 및 예측하는 실험을 진행하였다. USE, KorBERT 두 가지의 모델을 2010~2019년 재결서 데이터로 학습하여 선박사고의 사고 유형, 적용되는 법령, 형량을 분류 및 예측하였으며 그에 따른 정확도를 비교한 결과, KorBERT 문장 표상을 사용한 분류 모델이 가장 정확도가 높음을 확인했다.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.10a
• /
• pp.21-22
• /
• 2021

Early academic warning is considered as an inherent problem in education data mining. Early and timely concern and guidance can save a student's university career. It is widely assumed as a multi-class classification system in view of machine learning. Therefore, An accurate and precise methodical solution is a complicated task to accomplish. For this issue, we present a hybrid model employing rough set theory with a back-propagation neural network to ameliorate the predictive capability of the system with an illustrative example. The experimental results show that it is an effective early academic warning model with an escalating improvement in predictive accuracy.

• Journal of Information Processing Systems
• /
• v.20 no.2
• /
• pp.226-238
• /
• 2024

Recently, multi-access edge computing (MEC) has emerged as a promising technology to alleviate the computing burden of vehicular terminals and efficiently facilitate vehicular applications. The vehicle can improve the quality of experience of applications by offloading their tasks to MEC servers. However, channel conditions are time-varying due to channel interference among vehicles, and path loss is time-varying due to the mobility of vehicles. The task arrival of vehicles is also stochastic. Therefore, it is difficult to determine an optimal offloading with resource allocation decision in the dynamic MEC system because offloading is affected by wireless data transmission. In this paper, we study computation offloading with resource allocation in the dynamic MEC system. The objective is to minimize power consumption and maximize throughput while meeting the delay constraints of tasks. Therefore, it allocates resources for local execution and transmission power for offloading. We define the problem as a Markov decision process, and propose an offloading method using deep reinforcement learning named deep deterministic policy gradient. Simulation shows that, compared with existing methods, the proposed method outperforms in terms of throughput and satisfaction of delay constraints.