Region Merging Algorithm Based On Multi-measure Fusion For SAR Image Segmentation

Synthetic Aperture Radar(SAR)imaging technology has the unique advantage of allweather and all-time and thus has been widely used in remote sensing.With the development of the SAR imaging technology,the quality of SAR images has increasingly improved.However,due to the coherent imaging mode and speckle,the interpretation of SAR images is more difficult than that of optical images.As one of the important technologies of SAR image interpretation,SAR image segmentation is of importance.It is necessary to develop special segmentation methods for SAR images.In recent years,there have been an endless stream of SAR image segmentation methods.Due to the high complexity of real scene SAR images,no algorithm can segment SAR images with a mass of texture area accurately and efficiently.In this thesis,the segmentation method based on multi-scale Bhattacharyya distance is used for the initial segmentation of SAR images with textures.A new similarity measure defined in the frequency domain is proposed to characterize the difference between two textural regions.Based on these,a region merging SAR image segmentation algorithm using the multi-measure fusion cost function is given.The content of this thesis is arranged as follows:The first chapter introduces the research background of the thesis and the organization of the content.In the second chapter,the characteristics of SAR images and other related knowledge are firstly reviewed.Then,the SAR imaging mechanism and speckle model are introduced in details.Thirdly,the frequency domain characteristics of textures in SAR images are discussed.At last,several existing SAR image segmentation methods are reviewed.In the third chapter,an initial segmentation method based on multi-scale Bhattacharyya distance is introduced to embed into the existing Region Cost with Boundary Length Penalty(RCBLP)measure.It is improved by replacing the mean value using the Bhattacharyya distance between two regions.Furthermore,a measure is proposed to reflect the difference between two textural regions,which is calculated in the frequency domain by two different manners.Moreover,the textural similarity measure is fused into the improved RCBLP to construct a new region merging cost.Based on the new cost,a region merging SAR image segmentation algorithm is given,which is suitable for SAR images with a mass of texture areas.The new algorithm is compared with other algorithms on SAR images of real scenes.The advantages and disadvantages of individual algorithms are analyzed and quantitatively compared by edge-and region-based assessments.In the fourth chapter,a SAR image edge detection and region segmentation software demonstration system is built,which includes the commonly-used edge detection algorithms and segmentation algorithms of SAR images.The system can demonstrate various functions of the edge detection and region segmentation,including image selection and cropping,selection of edge detectors,parameter setting,hysteresis thresholding setting,edge map display,dynamic display of area merging process,and multi-staged display of results,etc.In the fifth chapter,the conclusion of the thesis is given and some future works are discussed.