類神經網路主成份分析器應用於腦波時空特性的分析和影像改良

Abstract

在訊號和影像的特徵抽取上，主成份分析是一個廣泛被使用的方法。在本 篇論文中我們介紹另一種方法，非監督適應型類神經網路方法來解決傳統 上用數學PCA處理的問題。本文將討論三個部份： (一)此類神經網路可用 來粹取多頻道腦波中具有空間關聯性的主要成份，並且我們用互關資訊來 計算其結果和數學PCA結果的相似程度。 (二)此網路能改善已退化影像的 品質，且我們將討論其結果和二維低頻濾波結果的不同處。我們發現PCA 網路優於濾波方法，因低頻濾波經常使影像變模糊。 (三)此網路稍加修 改後，即能消除具有空間關聯性的不想要的成份，此成份在多頻道間常佔 有相當重的比例。其結果我們將與數學分析方法的結果做個比較。 Principal component analysis (PCA) is a widely used method in feature extraction in signal and image processing. In this thesis, we present an alternative method, an unsupervised adaptive neural network method, to solve the problems which were conventionally studied by mathematic method (PCA). Firstly, we investigate the feasibility of applying adaptive neural network to extraction of spatially correlated features from multichannel EEGs. Mutual information function is used to quantitate the similarity in extracted components between network model and matrix algebra method. Secondly, we demonstrate that this network has the potential to enhance the quality of degraded images. In order to make a comparison, we also process the images using a 2-D lowpass filter. PCA network is superior to filtering method which usually blurs the images. Finally, we show that it is capable of removing spatially correlated unwanted component, which appears prominently among channels. In addition, we make a comparison between results of this spatial filtering network and matrix algebra method.