使用卷積神經網路之眼寫符號圖像辨識法;Pictorial Method with Convolutional Neural Network for Eye-writing Symbol Recognition

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/81561

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/81561

Title:	使用卷積神經網路之眼寫符號圖像辨識法;Pictorial Method with Convolutional Neural Network for Eye-writing Symbol Recognition
Authors:	張右新;Chang, Yu-Hsin
Contributors:	電機工程學系
Keywords:	眼電圖;卷積神經網路;眼動訊號眨眼偵測;眼寫系統;Electro-oculography(EOG);Convolutional Neural Network (CNN);EOG blinking detection;Eye-writing
Date:	2019-08-22
Issue Date:	2019-09-03 16:02:44 (UTC+8)
Publisher:	國立中央大學
Abstract:	Amyotrophic lateral sclerosis (ALS)是一種神經病變疾病，也被稱作Lou Gehrig‘s disease這種疾病特徵是會使腦中的運動神經不斷退化，也就是我們俗稱的漸凍人。患有ALS的病人四肢及軀幹的肌肉會逐漸地癱瘓麻痺甚至無力，同時也會慢慢喪失講話的功能，因此嚴重者會無法使用他們的肢體與口語溝通的能力。本研究的宗旨就是希望可以利用眼電圖法(ElectrooculoGraphy , EOG)做為眼動訊號的偵測來建構一套”眼寫系統”，並利用CNN網路來提高其辨識率。系統主要可以分成硬體和軟體兩大部分，而軟體的部分主要又可以分成：校準、雙眨眼偵測、符號分類三大功能，首先利用EOG訊號擬合出一個用來重建眼睛移動軌跡的轉移函數，有了這個轉移函數我們可以解決在時域上眼寫時符號的歪寫和比例大小問題，再者透過以訓練好的雙眨眼神經網路在時域上來進行雙眨眼偵測做為眼寫的開始，把截取之眼寫訊號轉換成29  29 pixel的二維圖像並強化圖像特徵，最後以卷積神經網路(CNN)來分類26個英文大寫字母、10個數字和4個命令符號(刪除、空白、換行、結束)。本論文利用CNN之圖像法來分類雙眨眼和字符大大的提高了其個人書寫的辨識率平均達到95%以上。 ;Amyotrophic lateral sclerosis (ALS) is a neuropathic disease. This disease is characterized by gradual degeneration of motor nerves in the brain, thus patients with ALS will experience progressive numbness and even weakness in the muscles of the extremities and trunks, and will slowly lose their speaking ability. Therefore, severe patients will not be able to move their body and communicate with others. The purpose of this study is to use the electro-oculogram method in an "eye-writing system" and use the CNN (convolutional neural network) network to improve its recognition rate. The main algorithm can be divided into three parts: calibration, double-blink detection, and symbol classification. First, the EOG signal at nine reference points is fitted to a transfer function to reconstruct the eye movement plane. With this transfer function, we can solve the problem of skewing and unequal scaling of the symbols in the time domain. Secondly, the trained CNN will detect double blinking from the EOG signal in the time domain to mark the beginning of the eye-writing of a symbol. Thirdly, the intercepted eye writing signal is converted into a 29 × 29 pixel two-dimensional image and the image features are enhanced. Finally, the CNN is used to classify 26 English letters,10 numbers and 4 command symbols (delete, blank, line feed, end). The CNN used in this study to classify double blinking and characters greatly improves the recognition of eye-writing symbols, with an average recognition rate over 95%..
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	206	View/Open

社群 sharing

Loading...