Font Size: a A A

A Study Of Multispectral Remote Image Segmentation

Posted on:2015-09-20Degree:MasterType:Thesis
Country:ChinaCandidate:C L LiFull Text:PDF
GTID:2308330464464584Subject:Optical Engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of the earth exploration remote sensing technology, the spatial resolution is more and more high, the satellite photograph has a clear terrain contour and rich details, it is very convenient for the rational utilization of the earth’s resources. A huge amounts of remote sensing images has been got from the earth satellites every year, and it has drawn much attention to study how to efficiently extract useful information from the large images. Image segmentation is one of the necessary links to extract terrain texture and is the foundation of the subsequent images application.Although there are many types of the image segmentation methods, and it has obtained certain research results, there are much details worth further studying due to the high resolution multispectral remote sensing image with clear geometric structure, more spectral bands, redundant information and large size of image data. In this paper, as the research object, the multispectral image were researched on the basis of image transform by using mathematical morphological methods and mean shift clustering algorithm. Meanwhile, problems of reducing image dimensions and reducing redundant information were analyzed in detail. Study results showed that compared with the traditional principal component analysis(PCA), Extended mathematical morphology method could keep much more details of the multispectral image characteristics and effectively describe the geometric structure of the terrain, and finally a hierarchical image segmentation results was obtained. The conducted experiments demonstrate the accuracy of the improved mean shift algorithm was increased and could get a better results.Firstly, the characteristics of high resolution multispectral image data was analyzed, focusing attention on redundant information and hierarchical segmentation.Despite the current commonly used PCA to remove the relevant information, the first principal component image lost part of the information; Different terrain category is of multi-scale, single segmentation result is difficult to accurately reflect the characteristics of the different scale of features. In this paper, on the basis of mathematical morphology,watershed transform and dynamics of contours could generate the multi-scale segmentation result and express the same object from different scales. Study results showed that compared with the traditional PCA method, the extended mathematical morphology could retain much more image details, and using dynamic contour can segment accurately and cause a multi-level segmentation result.Secondly according to the theory of extended morphology research results, it was applied into the mean shift clustering algorithm. Because of the many bands of multi-spectral images, there is large number of redundant information, it is greatly increase the difficulty of the calculation of the mean shift iterations and decrease the accuracy and efficiency of the mean shift algorithm. The extended morphology method and the watershed transform were used to control the redundant information of multi-spectral image, in the end using mean shift algorithm for image segmentation and merging small area to produce the final results.Finally for the application of remote sensing images, we explored methods to improve the segmentation of large-size images to reduce artificial border of block segmentation. The results show that the improved algorithm can improve the mean shift image segmentation large artificial boundaries. But it also requires further in-depth study and solves the problem of large-size image segmentation, really expanding the practical application of remote sensing images.
Keywords/Search Tags:Watershed transformation, large size segmentation, Mean Shift clustering
PDF Full Text Request
Related items