| Summary: | 碩士 === 國立臺灣大學 === 電機工程學研究所 === 92 === Hyperspectral images can provide spectral and geometrical information about target materials in the view. This makes it possible to remotely identify the target objects and their chemical compositions by discriminating spectra and shapes. But the amount of spectral image data generated from a hyperspectral imager is so huge that manual interpretation becomes very difficult or even impossible. Computerized automated interpretation becomes the only possible way to make full use of the information embedded in the hyperspectral image data. To achieve efficient automated image interpretation, reducing the data dimension by feature extraction is commonly invoked. For this purpose, the discrete wavelet transform (DWT) is used widely. But DWT-based feature extraction by collecting the approximation coefficients may lose some information about sharp absorption bands those are representative for a spectrum. Based on the full wavelet packet decomposition, this research has developed the Fusion following Wavelet Decomposition (FWD) to extract wavelet features from spectra. FWD is a novel method that can combine the information embedded in the low- and high-resolution channels of the wavelet decomposition, and as a result, detail features like sharp absorption bands can be preserved for recognition while reducing the data dimension. In the FWD feature extraction, significance of the wavelet features to represent the original spectrum is actually wavelet function dependent. The best wavelet function for a specific spectrum is found by optimizing the adaptive multi-channel wavelet decomposition.
An automated target recognition system based on FWD for spectral feature extraction and Learning Vector Quantization (LVQ) and Self-organizing Map (SOM) neural networks for automatic classification has been built in this research. First, the interested spectral classes in the hyperspectral image are identified and the image is transformed into a gray class map. Second, the edges of the target classes are extracted by contour tracing, and finally wavelet-based affine invariants of the contours are found to describe the shapes for automated recognition. Such a procedure of automated target recognition ensures to identify target objects by examining both their chemical and geometric properties.
|
|---|
