透過您的圖書館登入
IP:3.143.0.157
  • 期刊

Wavelet Networks for Hyperpsectral Image Classification

應用小波神經網路於高光譜影像分類

摘要


由於高光譜影像具有高維度資料的特性,在有限訓練樣本的情況下,傳統以統計參數爲基礎的影像分類法並無法直接適用。爲了降低維度以解決樣本量不足的問題,在分類前先進行光譜特徵萃取是最常用的方法。過去的研究指出以小波轉換爲基礎的小波特徵萃取能有效地萃取出對高光譜影像分類有用的特徵,然而在固定的尺度、平移參數下,作爲特徵的小波係數並無法依資料類別的特性而進行調整。本研究提出以小波神經網路爲基礎的分類器,其同時整合神經網路的訓練、分類以及小波特徵萃取,在訓練神經網路的過程中,同時動態調整小波轉換中的尺度、平移參數,並將神經網路中的權值最佳化,達到特徵萃取及分類的效果。由於訓練後小波神經網路的小波係數是依照各類別訓練樣本的特性而調整,作爲分類的特徵時,可提高各類別的分辨力。實驗結果顯示小波神經網路確實能提高小波特徵萃取的分類精度,且除了訓練樣本外,小波神經網路分類法幾乎不需要其他的先驗統計資訊,更增加了該分類法應用上的方便性。

並列摘要


Because of the high dimensionality of hyperspectral images, traditional statistics-based classifiers cannot be directly used for the hyperspectral image classification with limited training samples. The commonly used method to solve this problem is dimensionality reduction. Several research works have proven that wavelet transform provides an appropriate and effective tool for spectral feature extraction. Also, the researches about back-propagation neural networks for image classification have been investigated. In this study, wavelet networks is proposed for hyperspectral image classification. With wavelet networks, the important task of feature extraction is performed with wavelet transform, whereas classification is carried out by back-propagation neural networks. Both the position and the dilation parameters of the wavelets are optimized as well as the weights of the network during the training phase. The experiment results suggested that wavelet networks can be a useful and effective classifier for hyperspectral image classification.

延伸閱讀