• Proceedings of the Korea Information Processing Society Conference
• /
• 2017.11a
• /
• pp.857-860
• /
• 2017

현재 초보적인 능력을 가진 악기 연주자가 접근할 수 있는 하드웨어, 소프트웨어를 사용해 악기 연주법을 연습할 수 있는 수단은 전무하다. 따라서 본 논문은 악기 연주자가 연습을 하기 위해 사용할 수 있는 음 인식과 악보 정보의 처리, LSTM을 통한 자동 악보 생성의 복합적 기능을 가진 악기 보조 시스템을 제안한다. 또한 본 시스템은 기존의 FFT와 같은 일반적인 Pitch Detection 알고리즘보다 더 우월한 음 인식 성능을 보유한 Autocorrelation 전처리를 거친 LeNet-5 Convolutional Neural Network 모델을 사용하여 음 인식 성능을 높이는 기법을 제안한다. 이 음 인식 모델은 실험 결과 기존의 음 인식 기법보다 최대 약 5.4%의 성능 증가를 이루어냈다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.05a
• /
• pp.350-351
• /
• 2018

최근 의료 현장에 인공지능 기술의 도입이 가속화 되고 있다. 특히, 의료영상 분석 분야의 관련된 기 시스템 및 소프트웨어의 패러다임을 변화시키고 있다. 본 연구는 인공지능 기술을 적용하기 위한 학습의료영상 구성을 제안하고 이를 기반으로 X-ray 영상 중 손부위에 적용하여 오른손과 왼손을 판별하는 응용에 적용하였다. 그리고 Deep Learning Algorithm의 CNN을 개선하여 개발한 Advanced GoogLeNet를 적용하여 97%이상의 정확도를 보였다. 본 연구를 통해 얻어진 인공지능에 적용하기 위한 학습데이터 셋 구성과 개선된 알고리즘은 다양한 의료영상분석에 적용하고자 한다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.10a
• /
• pp.1004-1006
• /
• 2019

Different scale of information is an important component in computer vision systems. Recently, there are considerable researches on utilizing multi-scale information to solve the scale-invariant problems, such as GoogLeNet and FPN. In this paper, we introduce the notion of different scale information fusion (DSIF) and show that it has a significant effect on the performance of object recognition systems. We analyze the DSIF in several architecture designs, and the effect of nonlinear activations, dropout, sub-sampling and skip connections on it. This leads to clear suggestions for ways of the DSIF to choose.

• Journal of the Society of Naval Architects of Korea
• /
• v.59 no.2
• /
• pp.125-133
• /
• 2022

This paper proposes a fault detection method for a Unmanned Surface Vehicle (USV) with overactuated system. Current status information for fault detection is expressed as a scalogram image. The scalogram image is obtained by wavelet-transforming the USV's control input and sensor information. The fault detection scheme is based on Convolutional Neural Network (CNN) algorithm. The previously generated scalogram data was transferred learning to GoogLeNet algorithm. The data are generated as scalogram images in real time, and fault is detected through a learning model. The result of fault detection is very robust and highly accurate.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2020.07a
• /
• pp.559-560
• /
• 2020

우리는 패션 의류 이미지의 빠르고 정확한 분류를 달성하기 위해 최적화 된 동적 감쇠 학습률과 개선 된 모델 구조를 갖춘 딥 러닝 모델을 기반으로 하는 새로운 방법을 제안했습니다. 우리는 Fashion-MNIST 데이터 셋에 대해 제안 된 모델을 사용하여 실험을 수행하고 이를 CNN, LeNet, LSTM 및 BiLSTM의 방법과 비교했습니다.

• Journal of Institute of Control, Robotics and Systems
• /
• v.22 no.3
• /
• pp.241-246
• /
• 2016

Nowadays many people have an interest in facial expression and the behavior of people. These are human-robot interaction (HRI) researchers utilize digital image processing, pattern recognition and machine learning for their studies. Facial feature point detector algorithms are very important for face recognition, gaze tracking, expression, and emotion recognition. In this paper, a cascade facial feature point detector is used for finding facial feature points such as the eyes, nose and mouth. However, the detector has difficulty extracting the feature points from several images, because images have different conditions such as size, color, brightness, etc. Therefore, in this paper, we propose an algorithm using a modified cascade facial feature point detector using a convolutional neural network. The structure of the convolution neural network is based on LeNet-5 of Yann LeCun. For input data of the convolutional neural network, outputs from a cascade facial feature point detector that have color and gray images were used. The images were resized to $32{\times}32$. In addition, the gray images were made into the YUV format. The gray and color images are the basis for the convolution neural network. Then, we classified about 1,200 testing images that show subjects. This research found that the proposed method is more accurate than a cascade facial feature point detector, because the algorithm provides modified results from the cascade facial feature point detector.

• Nuclear Engineering and Technology
• /
• v.50 no.2
• /
• pp.223-228
• /
• 2018

AREVA NP (Courbevoie, Paris, France) is actively developing several enhanced accident-tolerant fuels cladding concepts ranging from near-term evolutionary (Cr-coated zirconium alloy cladding) to long-term revolutionary (SiC/SiC composite cladding) solutions, relying on its worldwide teams and partnerships, with programs and irradiations planned both in Europe and the United States. The most advanced and mature solution is a dense, adherent chromium coating on zirconium alloy cladding, which was initially developed along with the CEA and EDF in the French joint nuclear R&D program. The evaluation of the out-of-pile behavior of the Cr-coated cladding showed excellent results, suggesting enhanced reliability, enhanced operational flexibility, and improved economics in normal operating conditions. For example, because chromium is harder than zirconium, the Cr coating provides the cladding with a significantly improved wear resistance. Furthermore, Cr-coated samples exhibit extremely low corrosion kinetics in autoclave and prevents accelerated corrosion in harsh environments such as in water with 70 ppm Li leading to improved operational flexibility. Finally, AREVA NP has fabricated a physical vapor deposition prototype machine to coat full-length cladding tubes. This machine will be used for the manufacturing of full-length lead test rods in commercial reactors by 2019.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.18 no.3
• /
• pp.89-100
• /
• 2023

Recently, on-device artificial intelligence (AI) solutions using mobile devices and embedded edge devices have emerged in various fields, such as computer vision, to address network traffic burdens, low-energy operations, and security problems. Although vision transformer deep learning models have outperformed conventional convolutional neural network (CNN) models in computer vision, they require more computations and parameters than CNN models. Thus, they are not directly applicable to embedded edge devices with limited hardware resources. Many researchers have proposed various model compression methods or lightweight architectures for vision transformers; however, there are only a few studies evaluating the effects of model compression techniques of vision transformers on performance. Regarding this problem, this paper presents a performance evaluation of vision transformers on embedded platforms. We investigated the behaviors of three vision transformers: DeiT, LeViT, and MobileViT. Each model performance was evaluated by accuracy and inference time on edge devices using the ImageNet dataset. We assessed the effects of the quantization method applied to the models on latency enhancement and accuracy degradation by profiling the proportion of response time occupied by major operations. In addition, we evaluated the performance of each model on GPU and EdgeTPU-based edge devices. In our experimental results, LeViT showed the best performance in CPU-based edge devices, and DeiT-small showed the highest performance improvement in GPU-based edge devices. In addition, only MobileViT models showed performance improvement on EdgeTPU. Summarizing the analysis results through profiling, the degree of performance improvement of each vision transformer model was highly dependent on the proportion of parts that could be optimized in the target edge device. In summary, to apply vision transformers to on-device AI solutions, either proper operation composition and optimizations specific to target edge devices must be considered.

• Journal of the Korean Geographical Society
• /
• v.42 no.4
• /
• pp.469-487
• /
• 2007

Grain-Size Trend Analysis is the methodology to identify net sediment transport pathways, based on the assumption that the movement of sediment from the source to deposit leaves the identifiable spatial pattern of mean, sorting, and skewness of grain size. It can easily be implemented with low cost, so it has great potentials to contribute to geomorphological research, whereas it can also be used inadequately without recognition of its limitations. This research aims to compare three established methods of grain-size trend analysis to search for the adequate way of application, and also suggest the research tasks needed in improving this methodology 1D pathway method can corporate the field experience into analyzing the pathway, provide the useful information of depositional environments through X-distribution, and identify the long-term trend effectively. However, it has disadvantage of the dependence on subjective interpretation, and a relatively coarse temporal scale. Gao-Collins's 2D transport vector method has the objective procedure, has the capability to visualize the transport pattern in 2D format, and to identify the pattern at a finer temporal scale, whereas characteristic distance and semiquantitative filtering are controversial. Le Roux's alternative 2D transport vector method has two improvement of Gao-Collins's in that it expands the empirical rules, considers the gradient of each parameters as well as the order, and has the ability to identify the pattern at a finer temporal scale, while the basic concepts are arbitrary and complicated. The application of grain sire trend analysis requires the selection of adequate method and the design of proper sampling scheme, based on the field knowledge of researcher, the temporal scale of sediment transport pattern targeted, and information needed. Besides, the relationship between the depth of sample and representative temporal scale should be systematically investigated in improving this methodology.

• Asian-Australasian Journal of Animal Sciences
• /
• v.23 no.9
• /
• pp.1166-1173
• /
• 2010

Sixteen Mediterranean pluriparous buffaloes were subdivided into two uniform groups of eight animals. The average weight of the two groups at the start of the trial was 671.2 and 656.7 kg. The number of days from calving were 33.4 and 33.3, and the average milk production was 12.73 and 12.33 kg/d. The trial lasted for 114 days, and was divided into two sub-periods of 58 and 56 days. The two diets, administered ad libitum, had the same forage/concentrate ratio (53/47) but in their formulation the percentage of the two forages varied. Diet 1: alfalfa hay = 10%, maize silage = 43%, concentrate 1 = 47% (6.63 MJ/kg DM of net energy; 179.5 g/kg DM of crude protein). Diet 2: alfalfa hay = 20%, maize silage = 33%, concentrate 2 = 47%, (5.99 MJ/kg DM of net energy; 155.4 g/kg DM of crude protein). For the overall trial period (33-146 days in milk), the intake of dry matter was 17.23 kg/d for Group 1 and 17.29 kg/d for Group 2 and corresponded to 2.50 and 2.58% (p<0.01) of live weight. There was no significant difference between the average weight (689.7 and 669.4 kg) and the body condition score (6.49 and 6.42) of the two groups of buffaloes. Group 1 produced a greater quantity of milk (11.89 vs. 10.90 kg/d, p<0.10) of better quality both for its higher fat content (82.32 vs. 77.29 g/kg, p<0.10) and its protein content (47.36 and 46.38 g/kg). The milk produced by the buffaloes receiving Diet 1 had a better clotting ability, lower values of r (15.98 and 16.42 min) and K20 (1.66 and 1.75 min) and a higher value of A30 (54.45 and 52.73 mm). Taking into consideration the two sub-periods, milk production was significantly different only in the first sub-period (33-90 DIM), in favour of Group 1 (13.08 vs. 11.56 kg/d, p<0.05), while the positive effect of Diet 1 was cancelled out (10.71 and 10.24 kg/d) in the second part of the trial (91-146 DIM).