The Segmentation Research Of Object-level Multi-feature-based Markov Random Fields In Remote Sensing Images

With the successful launch of a large number of high spatial resolution remote sensing satellites at home and abroad in recent years,the data amount of observation to earth is getting larger and larger.However,the speed of remote sensing data processing in China cannot meet the urgent need for developing high-resolution remote sensing systems.So the processing of high resolution remote-sensing data and the intelligent transformation of spatial information has become the important scientific problems in current research.In the process of analysis for high-resolution images,image segmentation is an important basic step.For the segmentation of remote sensing images,each remote sensing image can be regarded as composed of random variables,and the amount of data contained in each remote sensing image is extremely large.Therefore,the modeling method based on probability statistics attracts much attention.Among them,the Markov Random Field Method(MRF)is widely used.However,there are some deficiencies in the existing MRF-based methods:(1)The object-based Markov Random Field approach(OMRF)based on mixed-Gaussian model does not consider the feature correlation between objects,and the segmentation accuracy is also limited;(2)The transfer between the neighboring scales in the multi-resolution model cannot fully utilize the segmentation results of each scale.In response to the above problems,this paper mainly made the following research work:An object-based Gaussian-Markov random field with the shape parameters of the object region(OGMRF-SP)method is proposed,which makes it possible to fully consider the interaction effects of neighboring objects' features when performing probability for the observed image data.First of all,in order to describe the complex interaction between the neighboring objects' features,the OGMRF-SP method uses the region size and boundary information of objects as the linear coefficients to construct an object-based linear regression equation(OLRE)for each object region.Second,the classical pixel-level Gaussian-Markov model is extended to the object level to model the error term in the OLRE.Then,OGMRF-SP method achieves the final segmentation result through probabilistic principle reasoning.Finally,in the experimental part,the appropriate selection experiments were performed for the artificially set parameters of this model.By comparing our experimental results with other existing MRF-based segmentation results on different data sets,it was shown that the proposed OGMRF-SP method has Better accuracy and efficiency.The object-based bilateral information transfer multi-resolution Markov random field(OB-MRMRF)is studied to make full use of the segmentation results of each scale in the multi-resolution model.First,the wavelet tool is used to construct the multi-resolution scale image structure,and the initial segmentation results are given on each scale.Then,using the bilateral information transmission mechanism,the adjacent upper-lower segmentation results are projected to the middle layer as an initial segmentation for iterative solution,and then the transmission mechanism is gradually projected onto the original resolution layer.Finally,compared with some existing MRF-based segmentation methods on different data sets,it is proved that our proposed OB-MRMRF method has higher segmentation accuracy.