SAR Image Segmentation Based On Deconvolutional Mapping Inference Networks

SAR image segmentation is a key job in the process of SAR image understanding and interpretation, which has a tremendous impact on the subsequent target detection and identification. For SAR image segmentation, traditional methods often require experience in the extraction of artificial feature, expression of these features are often very limited. Especially for the gathered feature in SAR image which have complex structure, most of the time it is difficult to design an effective feature to characterize it. The method of deep learning is able to learn feature automatically, thus reducing the manual intervention. At the same time it has the ability to learn the complex structural feature. So it can be used for learning structural feature in SAR image. However, it is impossible to accomplish the task of the SAR image segmentation directly just by feature, and it needs inference based on feature. Firstly, this paper uses Deconvolutional Networks to learn the structural feature in SAR image, and then proposes Mapping Inference Networks, based on which to do further reasoning, so as to realize the segmentation of SAR image. The main work of this paper is as follows:(1) This paper uses Deconvolutional Networks to learn structural feature of gathering area in SAR image. According to region map of SAR image, the SAR image is divided into gathering area, homogeneous area and structural area, so the task of SAR image segmentation is promoted to the semantic level. Then the overall task of SAR image segmentation is broken down into the segmentation of gathering area, homogeneous area and structural area. Among them, the structure of gathering area is very complex, so it's very difficult to realize the segmentation of gathering area effectively. Gathering area is formed by some non-connected regions. Then the segmentation of gathering area is to combine these non-connected regions based on feature. For these non-connected regions in gathering area, this paper samples and trains Deconvolutional Networks respectively, and then learns the filter set which stands for the structural feature of the region.(2) This paper proposes Mapping Inference Networks, which can realize the inference of feature,so as to complete the segmentation of gathering area in SAR image.. For nonconnected regions in gathering area, after the learning of the filter which stands for structure feature, in order to achieve the segmentation, the next task is to compare these regions based on the similarity of structure features. Based on self-organizing feature map, this paper proposes Mapping Inference Networks, so as to make a comparison between the structure feature of these regions, and thus to further complete the segmentation task of gathering area.(3) This paper uses the method based on gray feature and spectral clustering to complete the segmentation of homogeneous area in SAR image. Segmentation method for each type of area to be synthesized, this paper proposes a segmentation method of SAR image based on Deconvolutional Mapping Inference Networks and spectral clustering. The main idea of the method is: for gathering area, the method based on Deconvolutional Mapping Inference Networks is used; for homogeneous region, the method based on gray feature and spectral clustering is used; for structural area, the methods based on watershed is used; finally, the segmentation results of gathering area, homogeneous area and structural area are integrated, and the final segmentation result of the SAR image is obtained.