Multispectral MWIR image classification using filters derived from independent component analysis

In this research, a new data-dependent filtering based multispectral texture classification technique is presented. Multispectral basis functions were derived by applying Independent Component Analysis (ICA) on multispectral, mid-wave infrared (MWIR) images. These basis functions were then used as combined spatial-spectral filters to extract features from multispectral MWIR images. The process of fusing the features extracted from each band image was controlled by the phase relationship among the basis functions of the corresponding bands. The study of the multispectral basis functions revealed that phase opponencies between basis functions of different MWIR bands were similar to those seen among ICA basis functions derived from visible images. Multispectral Principal Component Analysis (PCA) basis functions from multispectral MWIR images were also extracted and used as filters for feature extraction. The quality of the features extracted using ICA-based filters, PCA-based filters, and multiscale opponent Gabor filtering are compared by measuring the classification rate of each technique. The ICA based filters had slightly classification performance than PCA and Gabor filters with no-added noise in the test images. However, ICA-based filters had far better classification rates than PCA-based filters and Gabor filters for noisy test images. The superior performance of ICA filters can be attributed to the fact that many ICA basis functions not only showed joint spatial-frequency resolution with properties similar to (though not the same as) Gabor filters but also had slightly more complex (richer) frequency response that were learned from the data.