結合多尺度主成分分析法與支持向量機在想像彩色圖像與中文文字之腦電波差異分析

Analysis the difference EEG of imaging color pictures and Chinese words via multi-scale principal component analysis and support vector machine

劉昀松(Liu yun-sung)

指導教授 ： 葉榮木 ； 蔡俊明

國立臺灣師範大學/科技學院/機電科技研究所/碩士(2010年)

摘要

大腦人機介面大多以想像左手、右手、腳與吐舌等體感動作為主，然而，以想像彩色圖像和中文文字作為大腦人機介面的媒介，目前少有相關研究團隊從事這方面研究，所以本研究在此試驗性地探討想像彩色圖像和中文文字的腦電波差異，並設計適合這種認知實驗的提示，此為本研究的主要動機和貢獻。
腦電波記錄時會包含著許多雜訊，例如:市電干擾、生理雜訊…等，而且腦電波訊號具有時變、非穩態等特性，所以腦電波的「特徵擷取」以及「分類」這兩個方面一直是主要的研究課題。本研究在「特徵擷取」方面，是採用多尺度主成分分析法來分析腦電波訊號，本分析方法分為兩個階段：第一階段是利用離散小波轉換中的多貝西小波，來分解腦電波訊號變成數個子頻帶，以增加特徵數量；第二階段是用主成分分析法，來得到最顯著的特徵值及其對應的特徵向量。本研究在「分類」方面，選擇具有高度辨識率的支持向量機，使用此來分類本研究所使用的彩色圖像與中文文字兩類想像內容。實驗受測者總共有13位，實驗結果發現以受測者S4為例，分類正確辨識率可高達88.89%，而平均正確準確率可達到72.65%。本研究更進一步探討圖像-動物類和文字-非動物類這兩種差異提示想像，實驗結果發現在圖像-動物類的平均分類正確率為74.08%，而文字-非動物類的平均分類正確率則達到84.21%，顯示對大部份受測者而言，文字-非動物類的平均分類正確率比圖像-動物類的平均分類正確率高。
本研究還有歸納出P300及N400在彩色圖像及中文文字提示的電位變化關係，N400事件相關電位在370ms~520ms間會出現一個負向的波峰，在負波峰後通常會有一個低頻的負波，這是腦電波感受到新奇刺激的反應，以及大腦空間能量分佈圖能量差異較大的頻帶位於Alpha2頻帶(11-14Hz)及Beta1頻帶(14-25Hz)，也就是常見的Beta頻帶。

關鍵字

腦電波 ； 大腦人機介面 ； 多尺度主成分分析法 ； 多貝西小波 ； 支持向量機

並列摘要

Brain computer interface (BCI) is mostly lead to complete by motor imagery which includes left hand, right hand, foot, and tongue. However, less of related groups put effort to study about BCI based on imaging of color pictures and Chinese words. In this thesis, the pilot study discusses the differences of electroencephalogram (EEG) of imaging color pictures and Chinese words, and design suitable hints of cognitive tasks. The above description is main motivation and the designed hints are main contributions in this study.
EEG recordings contain various noises, for instance, the disturbance from electric supply, physiological artifacts, etc. At the same times, EEG is time-variant and non-stationary. According to these reasons, “feature extraction” and “classification” of EEG are the main issues in BCI regions. In “feature extraction”, this study adopted multi-scale principle component analysis (multi-scale PCA) to analyze EEG signals. The whole method in this study is divided into two steps: In the first step, a kind of discrete wavelet transform, Daubechies wavelet, is used to decompose EEG signals into several sub-bands and increase the number of feature. In the second step, multi-scale PCA is used to extract most distinguished eigen values and eigen vectors. In “classification”, support vector machine (SVM) which usually has high accuracy is used to classify EEG of imaging color pictures and Chinese words in this study. There are thirteen subjects jointed this experiment. The performance in the experiment, one of subjects, S4 has highest accuracy of classification which can be reach to 88.89%, and averaged accuracy of all subjects is 72.65%. The more advanced research in this study is to discuss two different hints of imagery task, “pictures - animals” and “words - non-animal”. The experimental results show that the averaged accuracy of “pictures - animals” hint is 74.08% and “words - non-animal” hint is 84.21%. This could give a conclusion that “words - non-animal” hint has better performance on the accuracy rate of classification.
Finally, this study also generalized the relationship of EEG potentials between P300 and N400 during hints of “color pictures” and “Chinese words”. The phenomena in this study revealed that event-related potential with N400 was the negative peak during 370ms-520ms. In general, the wave with low frequency after the negative peak could exist. This is the response reduced from EEG when subjects receive strange stimulations. At the same times, the most distinguished difference between EEG of imaging “color pictures” and “Chinese words” with power spectrum is located at Alpha2 band (11-14 Hz) and Beta1 band(14-25 Hz). It is also called Beta band in the common.

並列關鍵字

Electroencephalography ； Brain–computer interface ； Multi-scale principal component analysis ； Daubechies wavelet ； Support vector machines

參考文獻

[1] Guest Editors, “Guest editorial the third international meeting on brain-computer interface technology: making a difference,” IEEE Trans. Rehabil. Eng., vol. 14, no. 2, pp. 126-127, 2006.

[2] F. Cincotti, L. Bianchi, G. Birch, C. Guger, J. Mellinger, R. Scherer, R. N. Schmidt, O. Y. Suárez, and G. Schalk, “BCI meeting 2005—workshop on technology: hardware and software,” IEEE Trans. Rehabil. Eng., vol. 14, no. 2, pp. 128-131, 2006.

[3] A. Kübler, V. K. Mushahwar, L. R. Hochberg, and J. P. Donoghue, “BCI meeting 2005—workshop on clinical issues and applications,” IEEE Trans. Rehabil. Eng., vol. 14, no. 2, pp. 131-134, 2006.

[4] A. Nijholt, B. Reuderink, and D. Plass-Oude Bos, “Turning shortcomings into challenges: Brain–computer interfaces for games,” Elsevier Entertainment Computing., vol. 1 pp. 85-94, 2009.

[5] A. Finke, A. Lenhardt, and H. Ritter, “The MindGame: A P300-based brain_computer interface game,” Elsevier Neural Networks., vol. 22, pp. 1329-1333, 2009.