CNN에 고밀도 근전도 신호 변환 방법을 적용한 손동작 인식 개선 기법

Alternative Title
Lee Joo Young
Author(s)
이주영
Alternative Author(s)
Lee Joo Young
Advisor
조위덕
Department
일반대학원 전자공학과
Publisher
The Graduate School, Ajou University
Publication Year
2020-02
Language
kor
Keyword
Convolutional Neural NetworkElectromyographyGesture RecognitionPhysiological Signal
Abstract
본 논문은 고밀도 근전도 신호에서 합성곱 신경망 기반의 동작인지에서 실행시간 대비 인식정확도를 고려한 시간에 따른 데이터 적층을 통한 이미지 변환 방법을 소개한다. 제안하는 방법은 2차원 고밀도 근전도 신호 이미지를 1차원으로 변환한 후 시간에 따라 적층하는 방법이다. 제안한 방법을 검증하고자, 공개되어있는 데이터 셋을 이용하여, 시간-창 길이와 중첩 길이에 따른 인식정확도와 실행시간을 측정했다. 이를 통해, 얻은 결과는, PC환경에서, 평균적으로 시간-창 길이에 따른 인식정확도 97.78 ± 0.16 %, 실행시간 0.21 ± 0.004 ms 였고, 중첩 길이에 따른 인식정확도 98.53 ± 0.38 %, 실행시간 0.25 ± 0.07 ms 였다. 제안된 방법은 종래 CNN에서 세 번째 차원에서 합성곱 연산이 불가한 구조에서, 기존 방법들에 비해 실행시간 대비 인식정확도 향상을 보였다.
Alternative Abstract
This paper introduces a time-stacking image transform method that considers execution time versus accuracy in gesture recognition of convolutional neural network (CNN) based analysis of high-density electromyography (HD-EMG) signals. The proposed method transforms two-dimensional HD-EMG signal images into one-dimensional representation and then stacks the images by time. To validate the proposed method, we measured the accuracy and execution time by time-window length and overlap length, using public data. The average recognition accuracy based on time-window length was 97.78 ± 0.16 % and the execution time was 0.21 ± 0.004 ms. The average recognition accuracy based on overlap length was 98.53 ± 0.38 % and the execution time was 0.25 ± 0.07 ms. In the structure of the CNN that does not perform convolution in the third dimension, the proposed method shows the improvement of execution time versus recognition accuracy compared with other CNN-based methods.
URI
https://dspace.ajou.ac.kr/handle/2018.oak/19487
Fulltext

Appears in Collections:
Graduate School of Ajou University > Department of Electronic Engineering > 3. Theses(Master)
Files in This Item:
There are no files associated with this item.
Export
RIS (EndNote)
XLS (Excel)
XML

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

Browse