Segmentation Algorithm Based On Fuzzy Markov Random Fields Of Remote Sensing Images

Remote-sensing image segmentation is the process to divide the image into regions with different features and extract the objective through segmentation,which will probably be used into the next processing step. Remote-sensing image segmentation is the important step from image processing to image analysis.In recent years, the unsupervised segmentation algorithms constrained by MRF as the priori probability model are widely used in the image segmentation. The experiments show that the model can significantly improve the quality of the image segmentation. As a result of influences of the environment and sensors, the characters of remote-sensing images are high gray level, sufficient texture and faint border,apply standard MRF to the speckle noises in segmentation can't give satisfying results. Aimed at the high levels of speckle noises and faint borders, in this paper, we apply the new unsupervised algorithm based on Fuzzy Markov random field to image segmentation. The Fuzzy Markov random field fuses the fuzzy set-based and the Markov model-based methods. The segmentation experiments of SAR images demonstrate that the proposed algorithm is correct. The main work and innovations of this paper is under following:(1) In this article, two main kinds of the algorithms, fuzzy set-based and Markov model-based algorithms, are discussed. We contrased the two algorithms in remote-sensing image segmentation.(2) In order to solve respective insufficiency of above two methods, the Fuzzy Markov random field is applied to remote-sensing image segmentation. Framework in segmentation algorithm based on Fuzzy Markov random field is proposed. Every pixel in image is not assigned a certain class but a fuzzy class. Based on the fuzzy random variables, the model combines the randomness and fuzziness in image segmentation. In terms of a soft segmentation algorithm, the more reasonably to receive the priori probability in accordance with the features of the image ,the more correct the priori probability is to be in the image segmentation.(3) When we devise the clique potential function in FMRF, we developed the Fuzzy MLL model based on traditional MLL. The clique potential function in Fuzzy model starts from the distance of two points and the classical clique potential function begins only from the differences of two points. So the clique potential function in Fuzzy model to describe the priori probability model is more carefully than the classical clique potential function. Then clique potential function in Fuzzy model to describe the subtle differences of the pixels is more carefully than the classical clique potential function in Gibbs random field.(4) 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 MAP-MRF framework, we extract the characters of gray and texture. We developed models respectively with their features. The experiments show that it is efficient to multi-level images.(5) There is a same problem in standard MRF and Fuzzy MRF which them highly depend on parameters estimated form image feature. There are less parameters used in this mothod, however, segmentation needs a ability to learn parameters and no longer to train datas to realize unsuperivsed image segmentation. In this article , we apply the EM algorithms to estimate unknown parameters and simulated annealing to search global optimal resolutions.(6) In the experiments, we compare our algorithm with fuzzy c-means clustering algorithm and FCM algorithm. The maximal error rate that is added by results of SAR image segmentation shows that FMRF is more efficient to deal with the edge overlapping. It reducs the speckle noises and improves the accuracy.(7) In experiments, author found that the fuzzy classification results in larger searching space. We can get satisfying result by simulated annealing that is a global optimal aglorithm, but the time of segmentation will be significantly longer. To improve efficiency of segmentation, the method is modified. The method combines SA with ICM. Although it would reduce the accuracy, the speed of segmentation enhanced nearly 3 times. So our method can adapt the different request.(8) We applied algorithms which are mentioned in this paper to SAR images segmentation with VC++ 6.0, and compared indicators of segmentation by different algorithms.