• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.05a
• /
• pp.464-466
• /
• 2019

CNN 은 합성곱 연산을 사용하는 인공신경망의 한 종류이다. 이러한 인공 신경망에서는 훈련 데이터에 대한 과도한 학습으로 인해 시험 데이터에 제대로 반응하지 못하는 오버피팅이 발생할 우려가 있다. 이를 해결하기 위해 DropOut 과 DropConnect 를 사용할 수 있다. 본 논문에서는 DropOut 과 DropConnect 를 통한 학습 정도를 실험을 통해서 비교해보고, 인공 신경망에서 이 방법의 효과를 살펴본다.

• Journal of Broadcast Engineering
• /
• v.26 no.6
• /
• pp.809-812
• /
• 2021

Air bubbles which may be generated during the PCB coating process can be a major cause of malfunction. so it is necessary to detect the bubbles in advance. In previous studies, candidates for bubbles were extracted using the brightness characteristics of bubbles, and the candidates were verified using CNN(Convolutional Neural Networks). In this paper, we propose a bubble detection method using a transfer learning-based CNN model. The VGGNet is adopted and sigmoid is used as a classification layer, and the last convolutional layer and classification layer are trained together when transfer learning is applied. The performance of the proposed method is F1-score 0.9044, which shows an improvement of about 0.17 compared to the previous study.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2023.05a
• /
• pp.120-120
• /
• 2023

본 연구에서는 CCTV 영상 내 빗줄기의 특성을 바탕으로 강우강도를 산정하기 위한 합성곱 신경망(CNNs, Convolutional Neural Networks) 기반 강우강도 산정 모형을 제안하였다. 중앙대학교 및 한국건설생활환경시험연구원 내 대형기후환경시험실에서 얻은 CCTV 영상들을 대상으로 연구를 수행하고, 우적계 등과 같은 지상 관측자료와 강우강도 산정 결과를 비교·검증하였다. 먼저, CCTV 영상 내 빗줄기의 미세한 변동 특성을 반영하기 위해 데이터 전처리 작업을 진행하였다. 이는 원본 영상으로부터 빗줄기 층을 분리해내는 과정, 빗줄기 층에서 빗물 입자를 분리해내는 과정, 그리고 빗물 입자를 인식하는 과정 등 총 세 단계로 구분된다. 합성곱 신경망 기반 강우강도 산정 모형 구축을 위해 영상 전처리가 완료된 데이터들을 입력값으로 설정하고, 촬영 시점에 대응되는 지상관측 자료를 출력값으로 고려하여 강우강도 산정모형을 훈련시켰다. CCTV 원자료 내 특정 영역에 편향되어 강우강도를 산정하는 과적합 현상의 발생을 방지하기 위해 원자료 내 5개의 관심 영역(ROI, Region of Interest)을 설정하였다. 추가로, CCTV의 해상도를 총 4개(2560×1440, 1920×1080, 1280×720, 720×480)로 구분함으로써 해상도 변화에 따른 학습 결과의 차이를 분석·평가하였다. 이는 기존 사례들과 비교했을 때, CCTV 영상을 기반으로 빗줄기의 거동 특성과 같은 물리적인 현상을 직간접적으로 고려하여 강우강도를 산정했다는 점과 더불어 머신러닝을 적용하여 강우 이미지가 갖는 본질적인 특징들을 파악했다는 측면에서, 추후 본 연구에서 제안한 모형의 활용 가치가 극대화될 수 있을 것으로 판단된다.

• Journal of IKEEE
• /
• v.22 no.4
• /
• pp.1115-1122
• /
• 2018

In this paper, we analyze the performance degradation of previous 1-Bit convolutional neural network method and introduce ways to mitigate it. Previous work applies 32-Bit operation to first and last layers. But our method applies 32-Bit operation to second layer too. We also show that nonlinear activation function can be removed after binarizing inputs and weights. In order to verify the method proposed in this paper, we experiment the object detection neural network for korean license plate detection. Our method results in 96.1% accuracy, but the existing method results in 74% accuracy.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.2
• /
• pp.130-138
• /
• 2021

This paper presents an effective structure by applying various normalization to Convolutional Neural Networks (CNN) for seismic event classification. Normalization techniques can not only improve the learning speed of neural networks, but also show robustness to noise. In this paper, we analyze the effect of input data normalization and hidden layer normalization on the deep learning model for seismic event classification. In addition an effective model is derived through various experiments according to the structure of the applied hidden layer. As a result of various experiments, the model that applied input data normalization and weight normalization to the first hidden layer showed the most stable performance improvement.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.4
• /
• pp.1-11
• /
• 2023

In this paper, instead of using recurrent neural network, we compare a classification performance of time series imaging algorithms using convolution neural network. There are traditional algorithms that imaging time series data (e.g. GAF(Gramian Angular Field), MTF(Markov Transition Field), RP(Recurrence Plot)) in TSC(Time Series Classification) community. Furthermore, we compare STFT(Short Time Fourier Transform) algorithm that can acquire spectrogram that visualize feature of voice data. We experiment CNN's performance by adjusting hyper parameters of imaging algorithms. When evaluate with GunPoint dataset in UCR archive, STFT(Short-Time Fourier transform) has higher accuracy than other algorithms. GAF has 98~99% accuracy either, but there is a disadvantage that size of image is massive.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.6
• /
• pp.759-767
• /
• 2020

This paper proposes a new Recurrent neural network (RNN) based spectrum sensing technique for cognitive radio communications. The proposed technique determines the existence of primary user's signal without any prior information of the primary users. The method performs high-speed sampling by considering the whole sensing bandwidth and then converts the signal into frequency spectrum via fast Fourier transform (FFT). This spectrum signal is cut in sensing channel bandwidth and entered into the RNN to determine the channel vacancy. The performance of the proposed technique is verified through computer simulations. According to the results, the proposed one is superior to more than 2 [dB] than the existing threshold-based technique and has similar performance to that of the existing Convolutional neural network (CNN) based method. In addition, experiments are carried out in indoor environments and the results show that the proposed technique performs more than 4 [dB] better than both the conventional threshold-based and the CNN based methods.

• The Journal of the Convergence on Culture Technology
• /
• v.8 no.6
• /
• pp.867-871
• /
• 2022

Amid the recent rapid trend change, the change in design has a great impact on the sales of fashion companies, so it is inevitable to be careful in choosing new designs. With the recent development of the artificial intelligence field, various machine learning is being used a lot in the fashion market to increase consumers' preferences. To contribute to increasing reliability in the development of new products by quantifying abstract concepts such as preferences, we generate new images that do not exist through three adversarial generative neural networks (GANs) and numerically compare abstract concepts of preferences using pre-trained convolution neural networks (CNNs). Deep convolutional generative adversarial networks (DCGAN), Progressive growing adversarial networks (PGGAN), and Dual Discriminator generative adversarial networks (DANs), which were trained to produce comparative, high-level, and high-level images. The degree of similarity measured was considered as a preference, and the experimental results showed that D2GAN showed a relatively high similarity compared to DCGAN and PGGAN.

• Journal of Broadcast Engineering
• /
• v.23 no.4
• /
• pp.484-494
• /
• 2018

The uniform motion blur removing algorithm of Cho et al. has the problem that the edge region of the image cannot be restored clearly. We propose the effective algorithm to overcome this problem by using shock filter that reconstructs a blurred step signal into a sharp edge, and convolutional neural network (CNN) that learns by extracting features from the image. Then uniform motion blur kernel is estimated from the latent sharp image to remove blur in the image. The proposed algorithm improved the disadvantages of the conventional algorithm by reconstructing the latent sharp image using shock filter and CNN. Through the experimental results, it was confirmed that the proposed algorithm shows excellent reconstruction performance in objective and subjective image quality than the conventional algorithm.

• Journal of Korea Society of Industrial Information Systems
• /
• v.25 no.3
• /
• pp.1-10
• /
• 2020

In this paper, we propose an algorithm for estimating blood pressure using ECG (Electrocardiogram) and PPG (Photoplethysmography) signals. To estimate the BP (Blood pressure), we generate a periodic input signal, remove the noise according to the differential and threshold methods, and then estimate the systolic and diastolic blood pressures based on the convolutional neural network. We used 49 patient data of 3.1GB in the MIMIC database. As a result, it was found that the prediction error (RMSE) of systolic BP was 5.80mmHg, and the prediction error of diastolic BP was 2.78mmHg. This result confirms that the performance of class A is satisfied with the existing BP monitor evaluation method proposed by the British High Blood Pressure Association.