Unsupervised Segmentation of Remote Sensing Images Based on Fuzzy Markov Random Field Model
ZHENG Wei,KANG Ge-wen,CHEN Wu-fan,LI Xiao-Wen
DOI: https://doi.org/10.3321/j.issn:1007-4619.2008.02.008
2008-01-01
Journal of Remote Sensing
Abstract:Remotesensing image segmentation is the process to divide the image into regions with different features and extract the object through segmentation,which will probably be used into the next processing step.Remotesensing image segmentatian is an important step from image processing to image analysis.In recent years,segmentation of remote sensing images plays the most important role in the interpretation of remote sensing images,but there is no reliable model to guide the remotesensing image segmentation.Markov Random Field(MRF) has been extensively applied in segmentation of images as priori probability model,and in fact,this method can be applied to improve the result of segmentation of images.Influenced by environment and the sensor,remote sensing images with complex texture,large brightness range and vague bridge boundary,can not be apt to use the standard MRF,and the result of segmentation of remote sensing images using standard MRF is not satisfactory.Therefore in remote sensing image segmentation,the algorithm using classic MRF is often inefficient.Aiming at ambiguity in the segmentation of remote sensing images,the Fuzzy Markow Random Field(FMRF) is put forward in this paper.Framework in segmentation algorithm based on FMRF is proposed.The method solves problems resulting from randomicity and fuzzability and gets the priori probability model reasonably,simultaneously,adds the gray feature and the texture feature into the segmentation model in order to get more accurate result.In the past,the techniques in fuzzy segmentation focused on the gray feature of image,so there are obvious defects with these techniques in multi-level classification.Under the Maximum a Posteriori Principle(MAP)MRF framework,we extract gray and texture features.We develope models respectively with their features.The experiments show that it is efficient to multi-level images.Finally,there are less parameters used in this method,however,segmentation needs ability to learn parameters and to realize unsupervised image segmentation. We apply the simulated annealing(SA) and expectation-maximization(EM) algorithm to estimate unknown parameters and get the global optimal solution.The segmentation experiments of SAR images demonstrate that the proposed algorithm is more efficient to distinguish interlaced edges and restrain the speckle noise than the algorithm of the Fuzzy C-Mean FCM and standard MRF.We apply algorithms that are mentioned in this paper to SAR images segmentation with VC++ 6.0,and compare indicators of segmentation by different algorithms.