• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.33 no.3
• /
• pp.375-381
• /
• 2023

Fully homomorphic encryption enables algebraic operations on encrypted data, and recently, methods for approximating non-algebraic operations such as the maximum function have been studied. However, precise approximation of max-pooling operations for four or more numbers have not been researched yet. In this study, we propose a precise max-pooling approximation method using the composition of approximate polynomials of the maximum function and theoretically analyze its precision. Experimental results show that the proposed approximate max-pooling has a small amortized runtime of less than 1ms and high precision that matches the theoretical analysis.

• Journal of Information Processing Systems
• /
• v.15 no.5
• /
• pp.1179-1191
• /
• 2019

We design an ingenious view-pooling method named learning-based multiple pooling fusion (LMPF), and apply it to multi-view convolutional neural network (MVCNN) for 3D model classification or retrieval. By this means, multi-view feature maps projected from a 3D model can be compiled as a simple and effective feature descriptor. The LMPF method fuses the max pooling method and the mean pooling method by learning a set of optimal weights. Compared with the hand-crafted approaches such as max pooling and mean pooling, the LMPF method can decrease the information loss effectively because of its "learning" ability. Experiments on ModelNet40 dataset and McGill dataset are presented and the results verify that LMPF can outperform those previous methods to a great extent.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.29-32
• /
• 2020

Due to the lack of improper image acquisition process, noise induction is an inevitable step. As a result, objective image quality assessment (IQA) plays an important role in estimating the visual quality of noisy image. Plenty of IQA methods have been proposed including traditional signal processing based methods as well as current deep learning based methods where the later one shows promising performance due to their complex representation ability. The deep learning based methods consists of several convolution layers and down sampling layers for feature extraction and fully connected layers for regression. Usually, the down sampling is performed by using max-pooling layer after each convolutional block. We reveal that this max-pooling causes information loss despite of knowing their importance. Consequently, we propose a better IQA method that replaces the max-pooling layers with strided convolutions to down sample the feature space and since the strided convolution layers have learnable parameters, they preserve optimal features and discard redundant information, thereby improve the prediction accuracy. The experimental results verify the effectiveness of the proposed method.

• 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.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.152-153
• /
• 2021

고성능 딥러닝 모델은 학습과 추론 과정에서 고비용의 전산 자원과 많은 연산량을 필요로 하여 이에 따른 개발 환경과 많은 학습 시간을 필요로 하여 개발 지연과 한계가 발생한다. 따라서 HW 또는 SW 개선을 통해 파라미터 수, 학습 시간, 추론시간, 요구 메모리를 줄이는 연구가 지속 되어 왔다. 본 논문은 EfficientNet에서 사용된 Linear Bottleneck을 변경하여 정확도는 소폭 감소 하지만 기존 모델의 파라미터를 55%로 줄이는 경량화 모델을 제안한다.

• Korean Journal of Remote Sensing
• /
• v.38 no.5_2
• /
• pp.793-810
• /
• 2022

Synthetic Aperture Radar (SAR) is considered to be suitable for near real-time inundation monitoring. The distinctly different intensity between water and land makes it adequate for waterbody detection, but the intrinsic speckle noise and variable intensity of SAR images decrease the accuracy of waterbody detection. In this study, we suggest two modules, named 'morphology module' and 'edge-enhanced module', which are the combinations of pooling layers and convolutional layers, improving the accuracy of waterbody detection. The morphology module is composed of min-pooling layers and max-pooling layers, which shows the effect of morphological transformation. The edge-enhanced module is composed of convolution layers, which has the fixed weights of the traditional edge detection algorithm. After comparing the accuracy of various versions of each module for U-Net, we found that the optimal combination is the case that the morphology module of min-pooling and successive layers of min-pooling and max-pooling, and the edge-enhanced module of Scharr filter were the inputs of conv9. This morphologic and edge-enhanced U-Net improved the F1-score by 9.81% than the original U-Net. Qualitative inspection showed that our model has capability of detecting small-sized waterbody and detailed edge of water, which are the distinct advancement of the model presented in this research, compared to the original U-Net.

• 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 Society of Computer Information Conference
• /
• 2021.07a
• /
• pp.65-66
• /
• 2021

본 논문에서는 Smart Factory의 자동 공정에서 결함의 분류를 실시간으로 시도하여 자동 공정 제어를 위한 결함 분류 딥러닝 기법을 제안하고, Pooling 종류에 따른 분류 성능을 비교한다. Smart Factory 구축에 있어서 CNN을 이용한 공정 제어를 통해 제품 생산에 있어서 생산량의 증가와 불량률의 감소를 이루어내는 것이 가능하다. Smart Factory는 자동화 공정이므로 결함의 분류 속도가 중요하지만, 생산량의 증가와 불량률의 감소를 위해서는 정확하게 결함의 종류를 분류하여 Smart Factory의 공정을 제어하는 것이 더욱 중요하다. 본 논문에서는 Pooling을 Max Pooling과 Averrage Pooling을 복합적으로 설정하였을 때 높은 성능을 보였다.

• 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.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.16 no.1
• /
• pp.41-49
• /
• 2023

A convolutional neural network (CNN) is a representative algorithm for implementing artificial neural networks. CNNs have improved on the issues of rapid increase in calculation amount and low object classification rates, which are associated with a conventional multi-layered fully-connected neural network (FNN). However, because of the rapid development of IT devices, the maximum resolution of images captured by current smartphone and tablet cameras has reached 108 million pixels (MP). Specifically, a traditional CNN algorithm requires a significant cost and time to learn and process simple, high-resolution images. Therefore, this study proposes an improved CNN algorithm for implementing an object classification learning model for simple, high-resolution images. The proposed method alters the adjacency matrix value of the pooling layer's max pooling operation for the CNN algorithm to reduce the high-resolution image learning model's creation time. This study implemented a learning model capable of processing 4, 8, and 12 MP high-resolution images for each altered matrix value. The performance evaluation result showed that the creation time of the learning model implemented with the proposed algorithm decreased by 36.26% for 12 MP images. Compared to the conventional model, the proposed learning model's object recognition accuracy and loss rate were less than 1%, which is within the acceptable error range. Practical verification is necessary through future studies by implementing a learning model with more varied image types and a larger amount of image data than those used in this study.