| Image segmentation is one of key pre-processing step that heavily influences theperformance of SAR image automatic interpretation. At present, it is one of popularresearch topic to search efficient, accurate segmentation method for SAR image. MRFmodel based approaches to image segmentation are used widely because it makes fulluse of the contextual information of image, however, these methods tend to be caught inlocal optimum, and are computationally expensive. In order to settle these problems, thethesis makes a deep analysis on MRF model-based approaches SAR to imagesegmentation. The main work includes the following aspects:(1) The statistical model of SAR image is incorporated into MRF model, thenEM algorithm is adopted to achieve the soft estimation of parameters of statisticalmodel and the soft decision of label image simultaneously. ROI of original image, suchas target, shadow are extratced efficiently, which is advantageous for automatic targetrecognition.(2) Because the classical nonlinear diffusion equation is inefficient and thesolution is not unique, it is modified to reduce the noise in cheaply computational cost.The numerical realization scheme of the new equation is put forward so as to reducespeckle of SAR image.(3) The clique potential function of classical MRF model is weighted by the localstatistic of original image, so both the label difference and intensity information aretaken into account. As a result, the boundary pixels are classified more accurately.(4) A framework for SAR image segmentation, which used the DS evidentialtheory within Markovian context, is introduced. Firstly, a basic probability assignmentbased on local statistic is defined to represent pixel evidence, then orthogonal sum isemployed to combine pixel evidence within Markov neighborhood, finally the labelimage is decided using specified rule such as the maximum a plausibility. Severalgroups of experiments demonstrate the better performance of the proposed method thanthe classical one. |
PDF Full Text Request