SAR Image Semantic Segmentation Based On Sketch Information And Feature Learning By Bayesian Network

With the rapid development of SAR imaging technology,the research of SAR image interpretation technology has become the focus,and SAR image segmentation is the premise and foundation for the study of SAR image interpretation technology.It is also an important part of the SAR image interpretation work.Because of the characteristics of SAR image,such as mixture,heterogeneity,high dimension and so on,the feature extraction of SAR image is difficult,which adds difficulty to the work of SAR image segmentation.The hierarchical visual semantic model proposed by the team divides the SAR image into hybrid,structural and homogeneous pixel subspaces.The model solves the heterogeneous problem of the SAR image well and lays a good foundation for the solution of the SAR image segmentation problem.On this basis,based on the information of the sketched line segment in the SAR image sketch map,a semantic segmentation method of SAR image based on sketch information and Bayesian network feature learning is proposed.The specific research results are as follows:(1)Put forward a segmentation method of hybrid pixel subspace based on sketch line segments direction information and feature learning.The sketch direction statistical vector is designed to represent the extremely inhomogeneous area in the hybrid pixel subspace of SAR image.The structure of the extremely inhomogeneous area is complex and the background information is abundant.Traditional methods are difficult to extract effective features for it.The SAR image sketch map is a sparse representation of the SAR image.The sketch line segment in the map contains rich semantic information.According to the characteristics of the sketch line segment,the design feature captures the obvious directional structure of the image.The primary clustering of hybrid pixel subspace is completed by clustering the artificial features.Combining the advantages of artificial features and learning features,the characteristics of the learning are obtained by combining the Bayesian learning network with the first classification results of artificial features,and the clustering method is used to complete the segmentation.According to the experiment,it is proved that the method is feasible.(2)A method based on clustering and feature similarity is proposed to analyze parameters of Bayesian learning network.The selection of network parameters directly affects the training ability and learning effect of network.The simulation experiment was completed through the different choices of network input layer neuron parameters and network hidden layer neuron parameters.The clustering method and feature similarity method based on variance and mean were used to analyze the parameters in combination with the experimental results.Select the best parameters,adjust the network structure,enhance the learning performance of network,and improve the accuracy of image segmentation.(3)A segmentation method based on sketch space structure and Bayesian learning network is proposed.According to the spatial position and length information of the sketch line segment,the constraints of the Bayesian learning network are designed to improve the network reconstruction effect and enhance the network learning performance to segment hybrid pixel subspace.The existing method is used to complete the segmentation of the homogeneous pixel subspace and the structural pixel subspace.The final SAR image segmentation results are obtained by combining the segmentation results of all pixel subspaces.The experimental verification shows that the research content of this paper is effective and feasible.