Research On Multiscale Analysis-Based Target Detection In Hyperspectral Imagery

With the development of the spectrometer technology, the research on hyperspectral imagery comes into a new stage emphasizing the effective process and utilization of the acquired data. Due to the high spectral resolution, narrow bandwidth and large amount of information, hyperspectral imagery can be use to distinguish and detect ground target with quite high diagnostic ability, so the research on target detection over hyperspectral imagery becomes a focus. However, the large quantity of data, high dimensionality and small target size make detection difficult and reduce the effectiveness of traditional detection methods severely. In this case, following aspects are researched in the thesis.Firstly, the characteristic of hyperspectral imagery and target detection method are researched. After analyzing the characteristic of spectral resolution, spatial correlation and band correlation, we take use of feature extraction techniques to reduce dimensionality, including principal component analysis and independent component analysis. Then, RX detection algorithm is introduced, for which we propose two improved methods, high-order moment-based ROI extraction and target principal component selection, to further increase the detection performance.Secondly, multiscale geometric analysis with application in target detection is studied and curvelet-based hyperspectral target detection algorithm is proposed. Current multiscale geometric analysis methods like curvelet transform possess good directivity, fast convergence and representing sparseness and perform better than wavelet transform in image processing. The thesis explains curvelet transform and its realization in detail; then emphasizes how to use curvelet transform to enhance hyperspectral target feature. The experimental results demonstrate the effectiveness and superiority of the proposed method.Finally, according to multiscale resolution analysis, we investigate the influence of different spatial, spectral and radiant resolutions to target detection. By introducing hyperspectral imaging spectrometer, we know the imaging principle, process and parameters of the used hyperspectral imagery. Then, multiscale spatial, spectral and radiant resolution analysis methods are brought forward, which can be applied to generate a series of hyperspectral imageries with various resolutions. At last, the detection results of these imageries show that a different kind of resolutions affects target detection in a distinct level. It is advisable to set appropriate resolutions in terms of requirement to make a trade-off between detection result and imaging cost.