| DC 欄位 |
值 |
語言 |
| DC.contributor | 電機工程學系 | zh_TW |
| DC.creator | 呂澤阜 | zh_TW |
| DC.creator | Tse-Fu Lu | en_US |
| dc.date.accessioned | 2019-8-22T07:39:07Z | |
| dc.date.available | 2019-8-22T07:39:07Z | |
| dc.date.issued | 2019 | |
| dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=105521073 | |
| dc.contributor.department | 電機工程學系 | zh_TW |
| DC.description | 國立中央大學 | zh_TW |
| DC.description | National Central University | en_US |
| dc.description.abstract | 本研究實現了一種基於卷積網路(Convolutional neural network, CNN) 的眼寫系統。 透過由眼電圖法(Electro-OculoGraphy, EOG)擷取眼睛移動產生的電訊號，採用End-to-End的架構，直接將訊號送入網路並且即時的輸出結果，無須再對眼電訊號進行繁雜的特徵抽取。本研究提出的系統除了硬體成本低廉，並且能夠在一般平價的品牌電腦上即時運行。
除了書寫之外，本研究提出的系統也能在各種控制場合做為人機介面，利用EOG訊號做為輸入指令，控制物件。本研究利用此系統實作一個範例，能夠直接依照著手寫筆劃書寫10個阿拉伯數字、26個英文大寫字母，以及自定義的空白(space)及退格(backspace)的眼寫系統，並且在測試時擁有95.7%的準確率。在文末也透過混淆矩陣(confusion matrix)來探討較容易使系統誤判的情況，並提供設計指令/符號的一些訣竅與建議。
此外，本研究還提出了即時的適應性眨眼偵測演算法用以偵測眨眼及計算眨眼次數來分隔符號、局部最大值偵測法用以縮放訊號來標準化訊號、及EOG訊號的兩種數據擴增方式以使數據集涵蓋更多可能的眼寫軌跡。
| zh_TW |
| dc.description.abstract | In this study, a Convolutional neural network (CNN) based eye-writing system is implemented. With the end-to-end structure, we acquire electric signal generated by eye moving through Electro-OculoGraphy (EOG), then feed the signal into the network directly without extracting complex features from the signal and network will output the symbol of recognition immediately. This system also has low hardware cost and it can run in most budget PC at real time.
Besides writing symbols or characters, this system also can be used to facilitate a Human-Machine Interface (HMI). In order to demonstrate the system proposed in this study, we used this system to implement an eye-writing system as an example which with 10 Arabic numerals and 26 English capital letters, all of which are written in accordance with handwritten strokes, as well as the space and backspace with specially designed writing patterns. This system attains a 95.7% accuracy in the test.
In the end of this paper, we discuss some conditions where it is easier to misjudge the system by the confusion matrix, and provide some tips and suggestions for instructions/symbols design.
In addition, this study also proposes a real-time adaptive blink detection algorithm which can detect and count eye blinks in the time domain with low computation, a local maximum based standardization algorithm and two method of data augmentation of EOG signal.
| en_US |
| DC.subject | 眼電圖 | zh_TW |
| DC.subject | 眼寫 | zh_TW |
| DC.subject | 眨眼偵測 | zh_TW |
| DC.subject | 深度學習 | zh_TW |
| DC.subject | 卷積網路 | zh_TW |
| DC.subject | 數據擴增 | zh_TW |
| DC.subject | EOG | en_US |
| DC.subject | eye-writing | en_US |
| DC.subject | blink detection | en_US |
| DC.subject | deep learning | en_US |
| DC.subject | CNN | en_US |
| DC.subject | data augmentation | en_US |
| DC.title | 資料擴增於即時眼寫辨識的卷積神經網路 | zh_TW |
| dc.language.iso | zh-TW | zh-TW |
| DC.title | Convolutional Neural Network with Data Augmentation for Real-time EOG Eye-writing Recognition | en_US |
| DC.type | 博碩士論文 | zh_TW |
| DC.type | thesis | en_US |
| DC.publisher | National Central University | en_US |