SAR Image Segmentation Based On Region Clustering

Synthetic aperture radar (SAR) has distinctive features of high-resolution imaging at all-time, all-weather conditions in contrast to traditional optical imaging. SAR imaging technology has been widely applied to wide range of fields such as military reconnaissance, satellite remote sensing, marine and land observing and other aspects for its many excellent features and other imaging abilities. Due to the coherent nature of the imaging process, images generated by SAR system are heavily contaminated by speckle noise which affects the segmentation result of SAR image.Gaussian mixture model (GMM) clustering algorithm is widely used in image segmentation in recent years. This algorithm is quite sensitive to speckle noise since spatial correlations between pixels are ignored. For this question, a region-clustering based segmentation algorithm for SAR image is presented in this dissertation by combining regional models with the GMM clustering algorithm. The main work of this dissertation is as follows:1. This dissertation reviews and summarizes the development history of SAR, SAR imaging systems and imaging principle (including the characteristics of SAR image data, the formation mechanism of speckle noise, etc.). A variety of SAR image segmentation algorithms are analyzed and compared.2. We present a region-based GMM clustering algorithm for SAR image segmentation. The watershed algorithm is firstly used to generate primitive homogeneous regions. By combining the spatial correlation of pixels, regional mean values are then calculated as the input samples of GMM clustering process. The impact of noise on the segmentation result can therefore be reduced in the space of regions instead of pixels.3. We present a novel initializing method for EM (Expectation- Maximization) algorithm. The self-feedback theory in cybernetics is referenced in this method and the parameters initialization of EM algorithm is simulated as a feedback system. The eventual estimate of parameters is guided by the initial estimate through feedback mechanism. As a result, the precision of parameter estimation is improved and a better clustering segmentation result is obtained while the iteration simplicity of EM algorithm is preserved. 4. Through a further research, we present an edge-preserving algorithm for SAR image segmentation. In this method, the SRAD filtering algorithm with the performance of edge preservation is applied to build a new region model, which promotes the integration of regions. Then the localization of the objective edge is more accurate and the over-segmentation within objective is much less.The results of segmentation experiments for both the synthetic images and the real SAR images demonstrate that the anti-noise ability and the segmentation accuracy of GMM clustering algorithm on region level has been substantially improved in contrast to the GMM algorithm on pixel level. So it has vast value in practical utility, particularly on the SAR images with strong speckle noise. The segmentation effect and efficiency of the algorithm with SRAD filtering are all improved greater than the former algorithm, which proves its practicability and validity in SAR image interpretation.