Study On Object Extraction Based On Multi-feature From Hyperspectral Image

This thesis was based on the property of syncretism of graph and spectrum for hyperspectral image. By excavating the inherence characteristic of hyperspectral remote sensing image and developing several effective analysis tools of image processing, high precision and robust method of object detecting and recognization was studied for typical objects and background in this paper.Firstly, the application and development of hyperspectral remote sensing image were reviewed. And then this paper pointed out and studied the existed problems of hyperspectral remote sensing image.Secondly, a matching and recognition method based on spectral subsection was researched. As well known, spectral matching technology is one of the key issues in the field of hyperspectral remote sensing image since it can be used for object recognition by comparing the spectral curves which provided the eradiating feature of ground objects. The spectral radiation feature is distributed to the entire spectrum and presented as several absorbing peaks and valleys with different scales. Based on this understanding, the study of extracting object radiation feature should be performed by taking multi-scale analysis technology which is helpful to extracting the spectral properties completely. In this thesis, the discrete wavelet transform is utilized for the application of spectral feature extraction. To this purpose, we took two order derivative of Gaussian function as wavelet basis and found the optimal inflexions based on multi-scale analysis, and presented a spectral subsection matching method based on the optimal inflexions.Thirdly, considering the shadow phenomena presented in hyperspectral images, a shadow detection method was described based on the density clustering of multiple bands and features by start on the spectrum analysis. The rough idea of this method covers (a) obtain the shadow region by segmenting the data with different features by utilizing the dynamic thresholding density clustering method and (b) a robust and effective shadow detection method was presented by a multiple evidences decision strategy for different shadow results obtained by different features.To remove the effect of shadow for the precision of image classification and reorganization, the existed shadow removing methods were discussed. Since the shadow information is sparseness, an adaptive shadow removal method was proposed based on tensor inpainting technology and radiation transmission correction technology. The procedure of this method includes that (a) conferring the statistics property of illumination and brightness by tensor analysis and voting technology and (b) adaptive enhancing the shadow regions by using radiation transmission model.Finally, we studied road detection method based on the property of syncretism of graph and spectrum for high spectral image. The first step of this method is to detect the underlying region of shadow by using spectral matching method, and then by integrating the geometric characteristic of road which was extracted by directional morphological operations, the roads were detected completely. To overcome the ruptures and discontinuities caused by the different disturbing factors and enhance the completeness of extracting road, the rupture roads were inpainted by extracting and inpainting the main direction of road based on tensor analysis.