Research And Implementation Of Remote Sensing Image Segmentation Algorithm

Image segmentation is the main step from image process to image analysis. On the one hand, image segmentation is the foundation of target expression, and has vital effect on feature survey. On the other hand, original image can be translated into more abstract and more compact format by image segmentation and target expression, feature extraction, parameter survey, and so on. This makes higher image analysis and image understanding possible.The characters of remote sensing image are higher gray level, tremendous image information, faint border and complex structure. And there is no reliable model to guide the remote sensing image segmentation. Therefore, the image segmentation technology is not used in the remote sensing field well. Considering the character of the remote sensing image, this thesis has studied on and improved the tranditional image segmentation algorithm. We make it fit for the character of the remote sensing image and apply it to remote sensing image segmentation. The work and contributions in this thesis are as follows.The current and mainstream image segmentation algorithms have been analyzed, classified, induced and summarized. Then we give the advantages and disadvantages of all kinds of methods. This provides definite bases for the people who choose different segmentation algorithms in the diverse application occasions and in the dissimilar data conditions.An effective method for image segmentation, Otsu algorithm based on threshold, is analyzed in this paper. As the Otsu algorithm is very slow in processing the multi-threshold segmentation, we optimize the Otsu algorithm with Nelder-Mead method.An effective method for edge detection, SUSAN althorithm, is analysed in this paper. Then we improve the edge detection precision of SUSAN althorithm, making it clear and accuracy to abstract the boundary of the landscape.About how to confirm seed points of region growth segmentation algorithm, we integrate it with watershed segmentation. Combining with the principle of the least heterogeneity area, we both remedy the over-segmentation of the watershed algorithm and solve the problem about comfirming seed points of the region growth segmentation.