Researches On Road Extraction Algorithm Based On GMM-MRF For High Resolution Remote Sensing Image

In remote sensing images, road is not only a kind of important basic geographic information, but also regarded as clues and reference for extraction of other ground objects. Correct road extraction is crucial to application of high resolution remote sensing images. However, high resolution remote sensing images have huge information with more and more road objects. There are some interrupts caused by trees or shadows of buildings. So, road extraction from high resolution remote sensing images has its scientific meaning and practical value.As a method of texture segmentation, the Gaussian mixture model and Markov random field (GMM-MRF) model could describe the priori information appropriately, such as the changing rate of the gray, the geometric features edge-based, and texton arrangement of an image, etc. It can build a priori distribution model through priori knowledge and can effectively characterize the spatial correlation of image. So GMM-MRF has a high value in image processing. This paper focuses on the application of GMM-MRF in road extraction from the remote sensing images.The main contents of this paper are as follows:1. A semi-automatic algorithm for road extraction from high resolution remote sensing images based on GMM-MRF and fuzzy connectedness. Firstly, the images are divided by GMM-MRF texture segmentation method into three classes. Then select a waypoint as a seed manually, and calculate fuzzy connectedness between each pixel and the seed point, and determine the initial target area by comparing fuzzy connectednesses. Secondly, the paper uses the method of shape features and mathematical morphology operations for further processing. Finally, the methods for evaluating the results of road extraction are done by mapping and quantitative indicators objectively. The results show that the proposed algorithm can extract the road area correctly.2. An automatic algorithm for road extraction based on covariance matrix. Firstly, the images are divided into3classes by the same segmentation method as1. Then calculate the mean of covariance matrix of each class. Nextly select the class that with the maximum mean as target, and treat the others as background. So the image realizes binaryzation, and we get the binaryzation image that contains road. Secondly, use the same post-processing method mentioned above to do the further processing of the extraction result. Finally, use the same evaluation method mentioned above to evaluate the post-processing result. The road extraction system for high resolution remote sensing images based on GMM-MRF texture segmentation model is developed with VC++6.0. The system implements the following functions: preprocessing of geometric correction, image graying and normalization etc.; the texture segmentation method based on GMM-MRF model; semi-automatic extraction from high resolution remote sensing images based on road GMM-MRF and fuzzy connectedness; an automatic algorithm for road extraction based on covariance matrix.