• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.9
• /
• pp.1158-1165
• /
• 2021

This paper proposes a CNN accelerator which is optimized Pooling layer operation incorporated in Multiplication And Accumulation(MAC) to reduce the memory size. For optimizing memory and data path circuit, the quantized 8bit integer weights are used instead of 32bit floating-point weights for pre-training of MNIST data set. To reduce chip area, the proposed CNN model is reduced by a convolutional layer, a 4*4 Max Pooling, and two fully connected layers. And all the operations use specific MAC with approximation adders and multipliers. 94% of internal memory size reduction is achieved by simultaneously performing the convolution and the pooling operation in the proposed architecture. The proposed accelerator chip is designed by using TSMC65nmGP CMOS process. That has about half size of our previous paper, 0.8*0.9 = 0.72mm². The presented CNN accelerator chip achieves 94% accuracy and 77us inference time per an MNIST image.

• 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.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.19 no.4
• /
• pp.161-166
• /
• 2018

In this paper, we propose a method for the random selection of pooling operations for the regularization and reduction of cross validation in convolutional neural networks. The pooling operation in convolutional neural networks is used to reduce the size of the feature map and for its shift invariant properties. In the existing pooling method, one pooling operation is applied in each pooling layer. Because this method fixes the convolution network, the network suffers from overfitting, which means that it excessively fits the models to the training samples. In addition, to find the best combination of pooling operations to maximize the performance, cross validation must be performed. To solve these problems, we introduce the probability concept into the pooling layers. The proposed method does not select one pooling operation in each pooling layer. Instead, we randomly select one pooling operation among multiple pooling operations in each pooling region during training, and for testing purposes, we use probabilistic weighting to produce the expected output. The proposed method can be seen as a technique in which many networks are approximately averaged using a different pooling operation in each pooling region. Therefore, this method avoids the overfitting problem, as well as reducing the amount of cross validation. The experimental results show that the proposed method can achieve better generalization performance and reduce the need for cross validation.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.87-90
• /
• 2020

GPU의 발전과 함께 성장한 딥러닝(Deep Learning)은 영상 분류 문제에서 최고의 성능을 보이고 있다. 그러나 합성곱 신경망 기반의 모델을 깊게 쌓음에 따라 신경망의 표현력이 좋아짐과 동시에 때로는 학습이 잘되지 않고 성능이 저하되는 등의 부작용도 등장했다. 성능 향상을 방해하는 주요 요인 중 하나는, 차원감소 목적에 따라 필연적으로 정보 손실을 겪어야 하는 풀링 계층에 있다. 따라서 특성맵(Feature map)의 차원감소를 통해 얻게 되는 비용적 이득과 모델의 분류 성능 사이의 취사선택(Trade-off)이 존재한다. 그리고 이로부터 자유로워지기 위한 다양한 연구와 기법이 존재하는데 Spectral Pooling도 이 중 하나이다. 본 논문에서는 이산 푸리에 변환(Discrete Fourier Transform, DFT)을 이용한 Spectral Pooling에 대한 소개와, 해당 풀링의 성질을 통상적으로 사용되고 있는 Max Pooling과의 성능 비교를 통해 분석한다. 또한 영상 내 고주파수 부분에서 특히 더 강건하지 못하다는 맥스 풀링의 고질적인 문제점을, Spectral Pooling과의 하이브리드(Hybrid) 구조를 통해 어떻게 극복해나갈 것인지 그 가능성을 중심으로 실험을 수행했다.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2018.05a
• /
• pp.466-467
• /
• 2018

Recently, artificial intelligence has been applied to various fields such as image recognition, image recognition speech recognition, and natural language processing, and interest in Deep Learning technology is increasing. Many researches on Convolutional Neural Network(CNN), which is one of the most representative algorithms among Deep Learning, have strong advantages in image recognition and classification and are widely used in various fields. In this paper, we propose a new network structure that transforms the general CNN structure. A typical CNN structure consists of a convolution layer, ReLU layer, and a pooling layer. Therefore in this paper, We intend to construct a new network by adding fully connected layer inside a general CNN structure. This modification is intended to increase the learning and accuracy of the convoluted image by including the generalization which is an advantage of the neural network.

• Smart Structures and Systems
• /
• v.32 no.3
• /
• pp.135-151
• /
• 2023

Using 3D asphalt pavement surface data, a deep and multiscale network named CrackNet-M is proposed in this paper for pixel-level crack detection for improvements in both accuracy and robustness. The CrackNet-M consists of four function-specific architectural modules: a central branch net (CBN), a crack map enhancement (CME) module, three pooling feature pyramids (PFP), and an output layer. The CBN maintains crack boundaries using no pooling reductions throughout all convolutional layers. The CME applies a pooling layer to enhance potential thin cracks for better continuity, consuming no data loss and attenuation when working jointly with CBN. The PFP modules implement direct down-sampling and pyramidal up-sampling with multiscale contexts specifically for the detection of thick cracks and exclusion of non-crack patterns. Finally, the output layer is optimized with a skip layer supervision technique proposed to further improve the network performance. Compared with traditional supervisions, the skip layer supervision brings about not only significant performance gains with respect to both accuracy and robustness but a faster convergence rate. CrackNet-M was trained on a total of 2,500 pixel-wise annotated 3D pavement images and finely scaled with another 200 images with full considerations on accuracy and efficiency. CrackNet-M can potentially achieve crack detection in real-time with a processing speed of 40 ms/image. The experimental results on 500 testing images demonstrate that CrackNet-M can effectively detect both thick and thin cracks from various pavement surfaces with a high level of Precision (94.28%), Recall (93.89%), and F-measure (94.04%). In addition, the proposed CrackNet-M compares favorably to other well-developed networks with respect to the detection of thin cracks as well as the removal of shoulder drop-offs.

• Journal of IKEEE
• /
• v.21 no.3
• /
• pp.244-247
• /
• 2017

In this paper, we implemented a convolution neural network using GP-GPU. After defining the structure, CNN performed inferencing using the GP-GPU with 256 threads, which was the previous study, using the weight obtained from the training. Training used Intel i7-4470 CPU and Matlab. Dataset used Daimler Pedestrian Dataset. The GP-GPU is controlled by the PC using PCIe and operates as an FPGA. We assigned a thread according to the depth and size of each layer. In the case of the pooling layer, we used over warpping pooling to perform additional operations on the horizontal and vertical regions. One inferencing takes about 12 ms.

• Journal of Communications and Networks
• /
• v.14 no.5
• /
• pp.495-500
• /
• 2012

A standard paradigm for the allocation of wireless resources in communication demands symmetry, whereby all users are assumed to be on equal footing and hence get equal shares of communication capabilities. However, there are situations in which "prime users" should be given higher priority, as for example in the transmission of emergency messages. In this paper, we examine a prioritization policy that can be implemented at the physical layer. In particular, we evaluate the performance of a prioritized transmission scheme based on spectrum pooling and on the assignment of higher signal-to-noise ratio channels to higher-priority users. This performance is compared to that of unprioritized (or "symmetric") schemes, and the impact of prioritization on the unprioritized users is discussed.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2018.11a
• /
• pp.97-98
• /
• 2018

최근 딥러닝 기술의 발전으로 영상 분류 및 영상 내 객체 검출뿐만 아니라 CNN 기반의 segmentation 기술도 개발되어 다른 요소까지 포함한 직사각형 영역의 검출 영역이 아닌 경계까지 고려한 분리가 가능하게 되었다. 더불어 사람 영역을 신체부위나 의류 부분과 같은 세부 영역으로 나누어 분리하는 human parsing 기술까지 연구되고 있다. Human parsing은 의류스타일 분석 및 검색, 사람의 행동 인식 및 추적과 같은 분야에도 응용될 수 있다. 본 논문에서는 Spatial pyramid pooling layer를 이용하여 영상 전체에 대한 공간적 분포 및 특성 정보를 고려한 human parsing 기법을 제안한다. Look into person(LIP) dataset을 이용하여 기존의 다른 segmentation 및 human parsing 기법과 제안하는 기법을 비교하여 제안하는 기법의 human parsing 결과가 보다 정교한 분리가 가능한 것을 확인하였다.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.6
• /
• pp.1-8
• /
• 2021

In this paper, we propose reinforced VGG style network structure for low performance embedded system to classify low resolution infrared image. The combination of reinforced VGG style network structure and global average pooling makes lower computational complexity and higher accuracy. The proposed method classify the synthesize image which have 9 class 3,723,328ea images made from OKTAL-SE tool. The reinforced VGG style network structure composed of 4 filters on input and 16 filters on output from max pooling layer shows about 34% lower computational complexity and about 2.4% higher accuracy then the first parameter minimized network structure made for embedded system composed of 8 filters on input and 8 filters on output from max pooling layer. Finally we get 96.1% accuracy model. Additionally we confirmed the about 31% lower inference lead time in ported C code.