Remote Sensing Image Segmentation Based On Anisotropic Markov Random Field

In recent years,with the continuous development of sensor technology,the resolution of remote sensing images is getting higher and higher,which has caused some scholars to have a keen interest in high-resolution images.Therefore,the processing of high-resolution remote sensing images has become an issue of great concern in the field of remote sensing.Remote sensing image segmentation occupies an important position in the field of remote sensing and is an important research issue in this field.The purpose of image segmentation is to divide the image into several regions.The internal properties of these regions are similar,but there are differences between the regions.There are currently many methods for correcting image segmentation,such as: clustering,threshold method,edge detection-based method,Markov random field(MRF)model,etc.Among them,the theoretical knowledge of the MRF model has gradually developed to maturity,and can describe the spatial information of the image in detail,so it is widely segmented.The detail features and structural features of the acquired images are increasing,and the model does not have a better segmentation effect on heterogeneous regions in the same category or homogeneous regions in different categories.As such,the problem of unbalanced data classification in the image also adds defects to the segmentation.Existing segmentation methods can often only guarantee the area,and cannot take into account the details(that is,the small area category is easily "swallowed" by the large area).In response to the above problems,this article mainly conducts the following research work:The object-based Markov Random Field with Anisotropic Penalty(OMRF-AP)method for remote sensing image segmentation is studied.Specifically,first,the input image is initialized and segmented to establish a region adjacency graph(GAG),define the neighborhood system,feature field and label field on the GAG,secondly,perform probabilistic modeling on the feature field and the segmented label field,and finally establish an anisotropic penalty matrix to describe the relationship between different categories and define the penalty Expected value of penalty information(EVPI)combines the anisotropic category interaction information with the posterior distribution information of the OMRF model,and iteratively updates the segmented EVPI items using the minimum posterior risk criterion to obtain the final image segmentation result.A Markov Random Field model based on Anisotropic Potential Function(APF-OMRF)is proposed.The model not only has good regionality,but also the anisotropic potential function passes setting different parameters can accurately describe the spatial interaction between different categories,so that the small area categories will not be simply negated.Specifically,first,the input image is initialized and over-segmented to establish a GAG.The neighborhood system,feature field,and label field are defined on the GAG.Second,the feature field is modeled for the likelihood function,the anisotropic potential function is defined,and the potential function is added to the joint probability distribution of the label field.Anisotropic probabilistic modeling is performed on the field.Finally,iterative segmentation is updated using the MAP criterion to obtain the final image segmentation result.In order to verify the effectiveness of the two models in this paper,we have compared with other methods.The experimental results on different high spatial resolution remote sensing images show that the proposed This model improves segmentation accuracy.