Research On Interpretation Techniques For High Resolution SAR Images Using Context Information

High resolution SAR images can provide strong support for the dailyintelligence and battlefield reconnaissance, with the rapid development of SARimaging technology and reconnaissance of rising frequency, the demand forautomatic interpretation technology for high resolution SAR images begin to rise. Indecimeter-level resolution, SAR images show two extra characteristics: scenecomplexity and intensity inhomogeneity, which make the traditional pixel-basedinterpretation techniques no longer applicable. Researching for high-resolution SARimage interpretation technique is critical to real time battlefield decision support andautomatic intelligence mining.This thesis starts from imitating the human visual interpretation system,proposed a high-resolution SAR image interpretation framework, whose core idea isutilizing context information in the image. It uses image segmentation as a base, andcase-based reasoning technology as a core part. Works including imagepreprocessing, image segmentation, target identification and target extaction arecompleted for localization vehicles from entire high resolution SAR images withcomplex scene. The main work can be summarized as follows:First, in order to make the original images with defects suitable for automaticprocessing, the pre-processing techniques for high-resolution SAR images arestudied. An automatic compensation mehtod of antenna beam roll-off based onpolynomial fitting is proposed, which is simple and robust. Then a speckle reductiontechnology based on non-local method is studied, which can utilizes spatial contextinformation in two different scales to preserve image detail while smoothing thespeckle noise in homogeneous regions. The preprocessing part enhances imagequality, and makes it possible for subsequent processing.Second, in order to get targets location and the object-level context information,high resolution SAR image segmentation technology is studied. Two classical levelset segmentation method—CV model and RSF model are analysed throughly, and animproved RSF model is proposed to solve the initialization problem of the RSFmodel. Then the improved RSF model is generalized to multiphase and applied tosegmentation tasks for high resolution SAR images, which consists of a backgroundsegmentation and a target segmentation. Experiments show that this method caneffectively segment high resolution SAR images without tuning parameters.Third, for high resolution SAR target discrimination problem, a case-basedreasoning expert system is proposed as a solution. Following the operation principleof human visual interpretation process, the rules of case matching and the data structure of case base is discussed and designed to built an expert system withcontinuous learning ability. Experiments show that the system can identify artificialtargets in complex environment effectively.Finally, with a practical application for vehicle extraction with high targetdensity, the dense target extraction technology for high resolution SAR images isstudied. In order to get vehicle pose and quantity information from positioning inthe target in dense target environment based on results of target discrimination. Avehicle "L" shaped structural feature extraction based on mathematical morphologyis proposed as the first stage, followed by template matching method to estimateposition and pose of the possible targets, then use the context constrain of vehiclegroups to screen the false targets. The method can get accurate result even in a verydense target environment.