• Smart Media Journal
• /
• v.12 no.7
• /
• pp.9-17
• /
• 2023

Plant diseases and pests affect the growth of various plants, so it is very important to identify pests at an early stage. Although many machine learning (ML) models have already been used for the inspection and classification of plant pests, advances in deep learning (DL), a subset of machine learning, have led to many advances in this field of research. In this study, disease and pest inspection of abnormal crops and maturity classification were performed for normal crops using YOLOX detector and MobileNet classifier. Through this method, various plant pest features can be effectively extracted. For the experiment, image datasets of various resolutions related to strawberries, peppers, and tomatoes were prepared and used for plant pest classification. According to the experimental results, it was confirmed that the average test accuracy was 84% and the maturity classification accuracy was 83.91% in images with complex background conditions. This model was able to effectively detect 6 diseases of 3 plants and classify the maturity of each plant in natural conditions.

• Journal of Broadcast Engineering
• /
• v.25 no.2
• /
• pp.200-207
• /
• 2020

Recently, deep neural networks such as CNN are showing excellent performance in various fields such as image classification, object recognition, visual quality enhancement, etc. However, as the model size and computational complexity of deep learning models for most applications increases, it is hard to apply neural networks to IoT and mobile environments. Therefore, neural network compression algorithms for reducing the model size while keeping the performance have been being studied. In this paper, we apply few compression methods to CNN models and evaluate their performances in the embedded environment. For evaluate the performance, the classification performance and inference time of the original CNN models and the compressed CNN models on the image inputted by the camera are evaluated in the embedded board equipped with QCS605, which is a customized AI chip. In this paper, a few CNN models of MobileNetV2, ResNet50, and VGG-16 are compressed by applying the methods of pruning and matrix decomposition. The experimental results show that the compressed models give not only the model size reduction of 1.3~11.2 times at a classification performance loss of less than 2% compared to the original model, but also the inference time reduction of 1.2~2.21 times, and the memory reduction of 1.2~3.8 times in the embedded board.

• International Journal of Computer Science & Network Security
• /
• v.21 no.7
• /
• pp.182-190
• /
• 2021

Deep learning is an advanced technology for large-scale data analysis, with numerous promising cases like image processing, object detection and significantly more. It becomes customarily to use transfer learning and fine-tune a pre-trained CNN model for most image recognition tasks. Having people taking photos and tag themselves provides a valuable resource of in-data. However, these tags and labels might be noisy as people who annotate these images might not be experts. This paper aims to explore the impact of noisy labels on fine-tuning pre-trained CNN models. Such effect is measured on a food recognition task using Food101 as a benchmark. Four pre-trained CNN models are included in this study: InceptionV3, VGG19, MobileNetV2 and DenseNet121. Symmetric label noise will be added with different ratios. In all cases, models based on DenseNet121 outperformed the other models. When noisy labels were introduced to the data, the performance of all models degraded almost linearly with the amount of added noise.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.3
• /
• pp.143-149
• /
• 2020

This paper presents the effect of the feature extraction methods used in the audio preprocessing on the classification performance of the Convolutional Neural Networks (CNN). We extract mel spectrogram, log mel spectrogram, Mel Frequency Cepstral Coefficient (MFCC), and delta MFCC from the UrbanSound8K dataset, which is widely used in environmental sound classification studies. Then we scale the data to 3 distributions. Using the data, we test four CNNs, VGG16, and MobileNetV2 networks for performance assessment according to the audio features and scaling. The highest recognition rate is achieved when using the unscaled log mel spectrum as the audio features. Although this result is not appropriate for all audio recognition problems but is useful for classifying the environmental sounds included in the Urbansound8K.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.2
• /
• pp.542-558
• /
• 2016

Vehicular Ad Hoc Networks (VANETs) are self-configuring networks where the nodes are vehicles equipped with wireless communication technologies. In such networks, limitation of signal coverage and fast topology changes impose difficulties to the proper functioning of the routing protocols. Traditional Mobile Ad Hoc Networks (MANET) routing protocols lose their performance, when communicating between vehicles, compromising information exchange. Obviously, most applications critically rely on routing protocols. Thus, in this work, we propose a methodology for investigating the performance of well-established protocols for MANETs in the VANET arena and, at the same time, we introduce a routing protocol, called Genetic Network Protocol (G-NET). It is based in part on Dynamic Source Routing Protocol (DSR) and on the use of Genetic Algorithms (GAs) for maintenance and route optimization. As G-NET update routes periodically, this work investigates its performance compared to DSR and Ad Hoc on demand Distance Vector (AODV). For more realistic simulation of vehicle movement in urban environments, an analysis was performed by using the VanetMobiSim mobility generator and the Network Simulator (NS-3). Experiments were conducted with different number of vehicles and the results show that, despite the increased routing overhead with respect to DSR, G-NET is better than AODV and provides comparable data delivery rate to the other protocols in the analyzed scenarios.

• International Journal of Advanced Culture Technology
• /
• v.9 no.4
• /
• pp.385-391
• /
• 2021

3D modeling of the human body is an integral part of computer graphics. Among them, several studies have been conducted on hair modeling, but there are generally few studies that effectively implement hair and face modeling simultaneously. This study has the originality of providing users with customized face modeling and hair modeling that is different from previous studies. For realistic hair styling, We design and realize hair segmentation using FCN, and we select the most appropriate model through comparing PSPNet, DeepLab V3+, and MobileNet. In this study, we use the open dataset named Figaro1k. Through the analysis of iteration and epoch parameters, we reach the optimized values of them. In addition, we experiment external parameters about the location of the camera, the color of the lighting, and the presence or absence of accessories. And the environmental analysis factors of the avatar maker were set and solutions to problems derived during the analysis process were presented.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.630-633
• /
• 2020

무의식적인 손-얼굴의 접촉으로 인한 감염의 문제점을 해결하기 위해, 얼굴 만지기 행동을 인식할 필요가 있다. 본 연구는 최근 각광을 받는 딥러닝 기술을 이용하여 비디오 영상에서 얼굴 만지기 행동 인식에 대한 연구이다. 우선, 비디오 영상에서 얼굴 만지기와 관련된 11 가지 행동에 대한 시, 공간적 특징을 컨볼루션 신경망을 통해 추출한다. 추출된 정보는 각 행동 레이블로 인코딩되어 비디오 영상에서 얼굴 만지기 행동을 분류한다. 또한, 3D, 2D 컨볼루션 신경망의 대표 네트워크인 I3D, MobileNet v3에 대해 비교 실험을 진행한다. 제안하는 시스템을 적용하여 인간의 행동을 분류하는 실험을 진행했을 때, 얼굴을 만지는 행동을 99%의 확률로 구분했다. 이 시스템을 이용하여 일반인이 무의식적인 얼굴 만지기 행동에 대해서 정량적으로 또는 적시적으로 인식을 하여, 안전한 위생 습관을 확립하여 감염의 확산방지에 도움을 줄수 있기를 바란다.

• Journal of the Korean Institute of Gas
• /
• v.8 no.2 s.23
• /
• pp.15-27
• /
• 2004

The purpose of this paper is to improve the performance of Polymer electrolyte fuel cell(PEMFC) by studying the channel dimension of bipolar plates using commercial CFD program 'Fluent'. Simulations are done ranging from 0.5 to 3.0mm for different size in order to find the channel size which shoves the highst hydrogen consumption. The results showed that the smaller channel width, land width, channel depth, the higher hydrogen consumption in anode. When channel width is increased, the pressure drop in channel is decreased because total channel length Is decreased, and when land width is increased, the net hydrogen consumption is decreased because hydrogen is diffused under the land width. It is also found that the influence of hydrogen consumption is larger at different channel width than it at different land width. The change of hydrogen consumption with different channel depth isn't as large as it with different channel width, but channel depth has to be small as can as it does because it has influence on the volume of bipolar plates. however the hydrogen utilization among the channel sizes more than 1.0mm which can be machined in reality is the most at channel width 1.0, land width 1.0, channel depth 0.5mm and considered as optimum channel size. The fuel cell combined with 2cm${\times}$2cm diagonal or serpentine type flow field and MEA(Membrane Electrode Assembly) is tested using 100W PEMFC test station to confirm that the channel size studied in simulation. The results showed that diagonal and serpentine flow field have similarly high OCV and current density of diagonal (low field is higher($2-40mA/m^2$) than that of serpentine flow field under 0.6 voltage, but the current density of serpentine type has higher performance($5-10mA/m^2$) than that of diagonal flow field under 0.7-0.8 voltage.

• Journal of Biomedical Engineering Research
• /
• v.45 no.3
• /
• pp.109-117
• /
• 2024

We developed a device that integrates digital otoscope and stethoscope for telemedicine. The integrated device was utilized for the collection of tympanic membrane images and cardiac auscultation data. Data accumulated on the platform server can support real-time diagnosis of heart and eardrum diseases using artificial intelligence. Public data from Kaggle were used for deep learning. After comparing with various deep learning models, the MobileNetV2 model showed superior performance in analyzing tympanic membrane data, and the VGG16 model excelled in analyzing cardiac data. The classification algorithm achieved an accuracy of 89.9% for eardrums data and 100% for heart sound data. These results demonstrate the possibility of diagnosing diseases without the limitations of time and space by using this platform.

• Nuclear Engineering and Technology
• /
• v.50 no.3
• /
• pp.356-367
• /
• 2018

A series of tests dedicated to station blackout (SBO) accident scenarios have been recently performed at the $Prim{\ddot{a}}rkreislauf-Versuchsanlage$ (primary coolant loop test facility; PKL) facility in the framework of the OECD/NEA PKL-3 project. These investigations address current safety issues related to beyond design basis accident transients with significant core heat up. This work presents a detailed analysis using the best estimate thermal-hydraulic code TRACE (v5.0 Patch4) of different SBO scenarios conducted at the PKL facility; failures of high- and low-pressure safety injection systems together with steam generator (SG) feedwater supply are considered, thus calling for adequate accident management actions and timely implementation of alternative emergency cooling procedures to prevent core meltdown. The presented analysis evaluates the capability of the applied TRACE model of the PKL facility to correctly capture the sequences of events in the different SBO scenarios, namely the SBO tests H2.1, H2.2 run 1 and H2.2 run 2, including symmetric or asymmetric secondary side depressurization, primary side depressurization, accumulator (ACC) injection in the cold legs and secondary side feeding with mobile pump and/or primary side emergency core coolant injection from the fuel pool cooling pump. This study is focused specifically on the prediction of the core exit temperature, which drives the execution of the most relevant accident management actions. This work presents, in particular, the key improvements made to the TRACE model that helped to improve the code predictions, including the modeling of dynamical heat losses, the nodalization of SGs' heat exchanger tubes and the ACCs. Another relevant aspect of this work is to evaluate how well the model simulations of the three different scenarios qualitatively and quantitatively capture the trends and results exhibited by the actual experiments. For instance, how the number of SGs considered for secondary side depressurization affects the heat transfer from primary side; how the discharge capacity of the pressurizer relief valve affects the dynamics of the transient; how ACC initial pressure and nitrogen release affect the grace time between ACC injection and subsequent core heat up; and how well the alternative feeding modes of the secondary and/or primary side with mobile injection pumps affect core quenching and ensure stable long-term core cooling under controlled boiling conditions.