• Journal of the Korea Society of Computer and Information
• /
• v.21 no.1
• /
• pp.1-8
• /
• 2016

Convolutional neural network (CNN) consists of a few pairs of both convolution layer and subsampling layer. Thus it has more hidden layers than multi-layer perceptron. With the increased layers, the size of convolution mask ultimately determines the total number of weights in CNN because the mask is shared among input images. It also is an important learning factor which makes or breaks CNN's learning. Therefore, this paper proposes the best method to choose the convolution size and the number of layers for learning CNN successfully. Through our face recognition with vast learning examples, we found that the best size of convolution mask is 5 by 5 and 7 by 7, regardless of the number of layers. In addition, the CNN with two pairs of both convolution and subsampling layer is found to make the best performance as if the multi-layer perceptron having two hidden layers does.

• Journal of Korea Multimedia Society
• /
• v.24 no.11
• /
• pp.1552-1559
• /
• 2021

Convolution neural networks (CNNs) show notable performance in image processing and are used as representative core models. CNNs extract and learn features from large amounts of train dataset. In general, it has a structure in which a convolution layer and a fully connected layer are stacked. The core of CNN is the convolution layer. The size of the kernel used for feature extraction and the number that affect the depth of the feature map determine the amount of weight parameters of the CNN that can be learned. These parameters are the main causes of increasing the computational complexity and memory usage of the entire neural network. The most computationally expensive components in CNNs are fully connected and spatial convolution computations. In this paper, we propose a Fourier Convolution Neural Network that performs the operation of the convolution layer in the Fourier domain. We work on modifying and improving the amount of computation by applying the fast fourier transform method. Using the MNIST dataset, the performance was similar to that of the general CNN in terms of accuracy. In terms of operation speed, 7.2% faster operation speed was achieved. An average of 19% faster speed was achieved in experiments using 1024x1024 images and various sizes of kernels.

• Journal of IKEEE
• /
• v.25 no.3
• /
• pp.493-500
• /
• 2021

In this paper, we propose the development of deep learning structure to improve quality of polygonal containers. The deep learning structure consists of a convolution layer, a bottleneck layer, a fully connect layer, and a softmax layer. The convolution layer is a layer that obtains a feature image by performing a convolution 3x3 operation on the input image or the feature image of the previous layer with several feature filters. The bottleneck layer selects only the optimal features among the features on the feature image extracted through the convolution layer, reduces the channel to a convolution 1x1 ReLU, and performs a convolution 3x3 ReLU. The global average pooling operation performed after going through the bottleneck layer reduces the size of the feature image by selecting only the optimal features among the features of the feature image extracted through the convolution layer. The fully connect layer outputs the output data through 6 fully connect layers. The softmax layer multiplies and multiplies the value between the value of the input layer node and the target node to be calculated, and converts it into a value between 0 and 1 through an activation function. After the learning is completed, the recognition process classifies non-circular glass bottles by performing image acquisition using a camera, measuring position detection, and non-circular glass bottle classification using deep learning as in the learning process. In order to evaluate the performance of the deep learning structure to improve quality of polygonal containers, as a result of an experiment at an authorized testing institute, it was calculated to be at the same level as the world's highest level with 99% good/defective discrimination accuracy. Inspection time averaged 1.7 seconds, which was calculated within the operating time standards of production processes using non-circular machine vision systems. Therefore, the effectiveness of the performance of the deep learning structure to improve quality of polygonal containers proposed in this paper was proven.

• ETRI Journal
• /
• v.39 no.6
• /
• pp.841-850
• /
• 2017

We investigate neural network image reconstruction for magnetic particle imaging. The network performance strongly depends on the convolution effects of the spectrum input data. The larger convolution effect appearing at a relatively smaller nanoparticle size obstructs the network training. The trained single-layer network reveals the weighting matrix consisting of a basis vector in the form of Chebyshev polynomials of the second kind. The weighting matrix corresponds to an inverse system matrix, where an incoherency of basis vectors due to low convolution effects, as well as a nonlinear activation function, plays a key role in retrieving the matrix elements. Test images are well reconstructed through trained networks having an inverse kernel matrix. We also confirm that a multi-layer network with one hidden layer improves the performance. Based on the results, a neural network architecture overcoming the low incoherence of the inverse kernel through the classification property is expected to become a better tool for image reconstruction.

• Journal of IKEEE
• /
• v.20 no.3
• /
• pp.303-306
• /
• 2016

In this paper, we propose a method to accelerate convolutional neural network by utilizing a GPGPU. Convolutional neural network is a sort of the neural network learning features of images. Convolutional neural network is suitable for the image processing required to learn a lot of data such as images. The convolutional layer of the conventional CNN required a large number of multiplications and it is difficult to operate in the real-time on the embedded environment. In this paper, we reduce the number of multiplications through Winograd convolution operation and perform parallel processing of the convolution by utilizing SIMT-based GPGPU. The experiment was conducted using ModelSim and TestDrive, and the experimental results showed that the processing time was improved by about 17%, compared to the conventional convolution.

• Journal of the Korean Society of Radiology
• /
• v.14 no.1
• /
• pp.39-44
• /
• 2020

The purpose of this study was learning about chest X-ray image classification and accuracy research through Deep Learning using big data technology with Convolution Neural Network. Normal 1,583 and Pneumonia 4,289 were used in chest X-ray images. The data were classified as train (88.8%), validation (0.2%) and test (11%). Constructed as Convolution Layer, Max pooling layer size 2×2, Flatten layer, and Image Data Generator. The number of filters, filter size, drop out, epoch, batch size, and loss function values were set when the Convolution layer were 3 and 4 respectively. The test data verification results showed that the predicted accuracy was 94.67% when the number of filters was 64-128-128-128, filter size 3×3, drop out 0.25, epoch 5, batch size 15, and loss function RMSprop was 4. In this study, the classification of chest X-ray Normal and Pneumonia was predictable with high accuracy, and it is believed to be of great help not only to chest X-ray images but also to other medical images.

• The Journal of the Convergence on Culture Technology
• /
• v.3 no.4
• /
• pp.165-169
• /
• 2017

CNN(Convolutional Nerual Network) is one of the algorithms that show superior performance in image recognition and classification among machine learning algorithms. CNN is simple, but it has a large amount of computation and it takes a lot of time. Consequently, in this paper we performed an parallel processing unit for the convolution layer, pooling layer and the fully connected layer, which consumes a lot of handling time in the process of CNN, through the SIMT(Single Instruction Multiple Thread)'s structure of GPGPU(General-Purpose computing on Graphics Processing Units).And we also expect to improve performance by reducing the number of memory accesses and directly using the output of convolution layer not storing it in pooling layer. In this paper, we use MNIST dataset to verify this experiment and confirm that the proposed CNN structure is 12.38% better than existing structure.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.11-12
• /
• 2023

GPGPU 환경에서의 ML 모델이 다양한 분야에 지속적으로 활용되면서, 이미지 분할(image segmentation) 연구가 활발하다. multi-GPU 환경에서 성능 최적화를 위하여 병렬화 기법들이 활용되고 있다. 본 연구에서는 multi-GPU 환경에서 U-Net 모델의 전체 수행 시간을 단축하기 위해 convolution 연산을 최적화하는 기법을 적용하는 실험을 진행하였고 shared memory, data parallelism 를 적용하여 82% 성능 향상을 보여주었다.

• The Journal of the Acoustical Society of Korea
• /
• v.15 no.5
• /
• pp.13-20
• /
• 1996

Recently, for implementation of neural networks extensive studies have been done especially VLSI technology has been regarded as the one of the most attractive means to implement neural networks. The main drawbacks of digital VLSI implementations are their large area and slow processing speed. In this paper to solve the speed and size problems we designed the efficient architecture using the binary convolution method for basic operation of neural cell, multiplication and addition. When it is used for implementing 3-layer network with 16 neural cell per layer that used neural cell based on binary convolution, clock of 50MHz and 26MCPS on 0.8${\mu}$ standard cell library has been achieved.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.15 no.1
• /
• pp.133-138
• /
• 2020

This paper introduces a method for improving accuracy using the first convolution layer, which is not used in most modified CNN(: Convolution Neural Networks). In CNN, such as GoogLeNet and DenseNet, the first convolution layer uses only the traditional methods(3×3 convolutional computation, batch normalization, and activation functions), replacing this with RGB-csb. In addition to the results of preceding studies that can improve accuracy by applying RGB values to feature maps, the accuracy is compared with existing CNN using a limited number of images. The method proposed in this paper shows that the smaller the number of images, the greater the learning accuracy deviation, the more unstable, but the higher the accuracy on average compared to the existing CNN. As the number of images increases, the difference in accuracy between the existing CNN and the proposed method decreases, and the proposed method does not seem to have a significant effect.