SAR Image Segmentation Based On Region Growing And Edge Penalty

In recent years,with the ability to image any targets on the earth, together and high resolution under nearly all weather conditions,Synthetic Aperture Radar(SAR) is widely applied in the field of military and economy. The method of abstracting the interested information in SAR image becomes more important. Rearching and developing the high accurate and low time-cost SAR image segmentation method has a great meaning.A new image segmentation algorithm named iterative region growing using semantics(IRGS)is proposed for standard image processing .It is an improvement over traditional Markov random field (MRF)-based approaches in that the edge strength information and a region growing technique are combined in formulating and searching objective functions. Considering a form of coherent speckle in SAR image,IRGS algorithm is not suitable for SAR image segmentation.In this thesis, a edge-preserving region model and a structure-preserving region model are proposed,which aims to solve the problem of the inaccurate edge location and over- segmentation. Rearching content is listed as follows:1.A edge-preserving region model combined with IRGS algorithm is proposed for SAR images segmentation.The improved algorithm uses speckle reduction anisotropic diffusion (SRAD) algorithm to reduce the impact of speckle noise. Then taking the watershed transform combined with Region adjacency graph (RAG) establishes the region expresstion of SAR image.2. A structure-preserving region model combined with IRGS algorithm for SAR images first apply The non-local means (NL-means) algorithm to remove speckle noise and restore image structures,then construct the region expresstion of SAR image by watershed transform and RAG.Both the proposed model combined with IRGS segmentation algorithm have been evaluated by using synthetic SAR images as well as a real SAR sea ice image. Relative to existing IRGS methods, testing results have demonstrated that the proposed new method substantially improves the segmentation performance and costs less time.