Feature combination for convolutional neural network number recognition
碩士 === 輔仁大學 === 資訊工程學系碩士班 === 107 === Feature capture is an indispensable part of image recognition. Usually, there are two directions for image recognition. One is a structured method, the other is to find similarities by a lot of statistics. The above two methods make mistakes easily when the char...
Main Authors: | , |
---|---|
Other Authors: | |
Format: | Others |
Language: | zh-TW |
Published: |
2019
|
Online Access: | http://ndltd.ncl.edu.tw/handle/m65s5e |
id |
ndltd-TW-107FJU00396029 |
---|---|
record_format |
oai_dc |
spelling |
ndltd-TW-107FJU003960292019-07-31T03:42:57Z http://ndltd.ncl.edu.tw/handle/m65s5e Feature combination for convolutional neural network number recognition 特徵組合之卷積神經網路的數字辨識 KUO,CHAO-WEI 郭兆偉 碩士 輔仁大學 資訊工程學系碩士班 107 Feature capture is an indispensable part of image recognition. Usually, there are two directions for image recognition. One is a structured method, the other is to find similarities by a lot of statistics. The above two methods make mistakes easily when the characters and images are slightly damaged.In this paper, we use convolution to extract features. We identify numbers just like the human brains. Even the numbers are complex, we can still recognize the characteristics of numbers from the perspective of human. The recognized features are fed to convolutional neural network to learn, more similar to the perspective of human identification.This paper extends the point, line and angle feature extractions proposed in [13], and adds the most important closed circle and mirror 2 features for identification. The original method cannot distinguish between number 2 and number 8 successfully. There is 56.8% probability to identify 8 as 2. By the method proposed in this paper, the discrimination rate of 2 and 8 is as high as 96.7%. After the feature of mirror 2 is used, the resolution is as high as 99%. KUO,WEN-YAN 郭文彥 2019 學位論文 ; thesis 48 zh-TW |
collection |
NDLTD |
language |
zh-TW |
format |
Others
|
sources |
NDLTD |
description |
碩士 === 輔仁大學 === 資訊工程學系碩士班 === 107 === Feature capture is an indispensable part of image recognition. Usually, there are two directions for image recognition. One is a structured method, the other is to find similarities by a lot of statistics. The above two methods make mistakes easily when the characters and images are slightly damaged.In this paper, we use convolution to extract features. We identify numbers just like the human brains. Even the numbers are complex, we can still recognize the characteristics of numbers from the perspective of human. The recognized features are fed to convolutional neural network to learn, more similar to the perspective of human identification.This paper extends the point, line and angle feature extractions proposed in [13], and adds the most important closed circle and mirror 2 features for identification. The original method cannot distinguish between number 2 and number 8 successfully. There is 56.8% probability to identify 8 as 2. By the method proposed in this paper, the discrimination rate of 2 and 8 is as high as 96.7%. After the feature of mirror 2 is used, the resolution is as high as 99%.
|
author2 |
KUO,WEN-YAN |
author_facet |
KUO,WEN-YAN KUO,CHAO-WEI 郭兆偉 |
author |
KUO,CHAO-WEI 郭兆偉 |
spellingShingle |
KUO,CHAO-WEI 郭兆偉 Feature combination for convolutional neural network number recognition |
author_sort |
KUO,CHAO-WEI |
title |
Feature combination for convolutional neural network number recognition |
title_short |
Feature combination for convolutional neural network number recognition |
title_full |
Feature combination for convolutional neural network number recognition |
title_fullStr |
Feature combination for convolutional neural network number recognition |
title_full_unstemmed |
Feature combination for convolutional neural network number recognition |
title_sort |
feature combination for convolutional neural network number recognition |
publishDate |
2019 |
url |
http://ndltd.ncl.edu.tw/handle/m65s5e |
work_keys_str_mv |
AT kuochaowei featurecombinationforconvolutionalneuralnetworknumberrecognition AT guōzhàowěi featurecombinationforconvolutionalneuralnetworknumberrecognition AT kuochaowei tèzhēngzǔhézhījuǎnjīshénjīngwǎnglùdeshùzìbiànshí AT guōzhàowěi tèzhēngzǔhézhījuǎnjīshénjīngwǎnglùdeshùzìbiànshí |
_version_ |
1719231418173751296 |