Research On SAR Image Segmentation Algorithm Based On MRF Model

Synthetic aperture radar(SAR) is a kind of two-dimensional imaging microwave radar,with the all-weather, all-day and all-round ability to detect and observe the surface of the earth Because of its convenient and strong ability of observation, SAR has been widely used in military and civil fields. So SAR technology has become the research hotspot to different countries and its imaging resolution is more and more high. But its development of technology is not consistent with the low processing rate and ability of interpretation of SAR image,which greatly limits the use rate of SAR image. While image segmentation is the prerequisite for the understanding and interpretation of SAR image, so the research on correct segmentation algorithm of SAR image is particular important. The main work of this paper is to study the SAR image segmentation algorithm based on markov random field(MRF) model,including spatial single-scale MRF model and multi-scale MRF model.Firstly, the imaging principle of SAR system and the mechanism of producing inherent speckle noise are studied in this paper. Then the statistics of SAR image gray distribution is analyzed and the fit of different distribution model for SAR image is obtained by simulation,The characteristics of SAR images is summarized through simulation results. Finally, the evaluation criterion of the quality of image segmentation is introduced to prepare for the follow-up study on SAR image segmentation algorithm.Secondly, the theory knowledge of spatial single-scale MRF model segmentation method,including the label field modeling, the observation modeling, the estimation method of parameters and the posterior segmentation method, has been systematic and detailedly studied.Based on the experimental simulation on several SAR images, the segmentation performance of the traditional k-mean clustering and the spatial single-scale MRF model segmentation algorithm are compared, and the error segmentation factors of MRF model for SAR image are analyzed and summarized.Thirdly, based on the characteristic of the SAR image with rich texture information, the textural feature is introduced into MRF model to improve its segmentation performance. The definition of texture and the extraction method of textural features are introduced, then the textural features of SAR image are analyzed. What’s more, the fuzzy theory is introduced into label field to improve the segmentation performance, because the SAR image is usually fuzzy.Experiments show that the segmentation result of the improved MRF model is better than spatial MRF model in regional consistency and edge preserving.Finally, multi-scale MRF model, including MSRF model and wavelet domain hierarchical Markov model, is studied. The wavelet domain hierarchical Markov model can better describe the global and detail information of SAR image by introducing inter-scale causal model and the inner-scale non-causal model. The simulation results show that the segmentation result of the wavelet domain hierarchical Markov model is better than spatial MRF model and MSRF model in regional consistency and edge preserving.