The Research On SAR Image Segmentation Technology

Synthetic aperture radar (SAR) is one of the representatives of microwave remote sensing. Since 1960, along with the SAR technology in the public, the role of earth observation gradually revealed. Since 2000, SAR technology kept a fast and steady development. Lots of new SAR data, brought new challenges for SAR information extraction and application.Faced with increasing SAR image resources, SAR interpretation technology development is still relatively lags behind. This makes the utilization of SAR image be greatly restricted. The segmentation of SAR image is the fundamental technology of SAR image interpretation. Under this background, we will focus on SAR image segmentation technology. We study the technology of active contour model (ACM) via level set and Markov random field (MRF) via graph cut. And we do some primary research on the extraction of semantic information from SAR images. We attempt to use some segmentation and classification algorithms for this purpose.The main research contents of this dissertation are as follows:(1) In the research of active contour model via level set, we first discuss the classical ACM based SAR image segmentation methods. And then we propose a new energy functional via level set. The energy functional is proposed to obtain a stationary global minimum by Euler-Lagrange equation. This functional can obtain a stationary global minimum so the level set function will converge to the same result no matter any kinds of initial contours. Moreover, we can easily set a terminal criterion for the level set based curve evolution.(2) In the research of Markov random field model via graph cut for SAR image segmentation, first, we analysis the present pixel-level MRF based SAR image segmentation and the main energy optimization algorithms. And then we focus on the SAR image segmentation based on MRF via graph cut.The energy function of the classical MRF based SAR image segmentation method contains two terms, which are data likelihood term and smoothness prior term. In this dissertation, we introduce a geometric prior term. Besides, we design a new multi-layer graph model to represent the new energy function. We can set different smoothness term parameters for different kinds of ground objects. The experimental results show that our proposed new method can get a better SAR image segmentation.(3) For now, how to extract the necessary information from massive SAR images is a practical problem for the utilization of SAR image to be resolved. One way to solve this problem is to submit the semantic information of SAR images to SAR image retrieval system. Then the retrieve system can extract the information more directly. A Good segmentation algorithm for SAR image semantic information extraction is very important. Actually, the extraction of SAR image semantic information will be greatly simplified, if good segmentation result is obtained.In this dissertation, we have done some preliminary work of SAR image semantic information extraction. We also present a basic framework for SAR image semantic information extraction. The applications of level set based ACM and graph cut based MRF for SAR image semantic information extraction are also presented.For the large scene SAR image, pixel-level based segmentation methods are not very suitable for semantic information extraction. In this case, we provide a solution which utilizes SAR image classification method via machine learning to obtain the semantic information.