Sar Image Segmentation Based On Level Set Methods

Synthetic Aperture Radar (SAR) plays an important role in modern remote sensing applications, because of its capability to operate at any time of day, in any weather conditions. With the rapid growth of available SAR data, the development of SAR image understanding techniques is more and more desired, and the efficient way for SAR image interpretation is now the urgent issue of remote sensing application. However, SAR imaging has the drawback of leading to images that are degraded by speckle noise which makes the images grainy and deteriorate the image processing performances. Therefore, the traditional digital image processing and analysis techniques that work successfully on optical images often do not perform as well as on SAR images.Segmentation is a crucial step in SAR image interpretation but it is also acknowledged as a difficult problem due to the presence of speckle. Recently, the level set method developed under partial differential equations (PDEs) framework provides a novel approach for image segmentation. However, this kind of approach is very difficult to be directly applied on SAR images because the information that describes SAR image features is not sufficiently used and the effect of speckle is not considered.In this thesis, the level set segmentation approaches that are applicable for SAR image are studied. Firstly, a level set approach which simultaneously considers the boundary and region features of SAR image is presented in this paper. The proposed approach defines a new edge-indicator based on SAR image texture and an energy function consist of region and boundary information. At the same time, we introduce a comparatively simple way to segment multi-region images by applying hierarchical splitting. Image segmentation is implemented by minimizing the energy criterion via variational level set approach. Experimental results show that this approach is efficient in SAR image segmentation without denosing process, and has the high accuracy of alignment to the real boundary.Additionally, we introduce a level set PDE based on the simplified Mumford-Shah model for PolSAR image segmentation proposed by Chan and Vese. Based on Cloude's polarimetric target decompositions theory, we can obtain two important parameters H,αand get the energy function for segmentation. It is the first time to address the PolSAR image segmentation by using level set method and polarization decomposition. However, the curve evolution would be seriously constrained by the parameters in PDE, and the increase of the number of multiple regions will raise the computation complexity and the time spent. So, finding the optimization algorithms will be the leading target in my future research.