SAR Image Segmentation Based On Information Of Sketch And Selection Of Image Block

The segmentation of SAR images is the premise and basis for the interpretation of SAR images.The quality of the segmentation results directly determines the quality of interpretation.SAR images contain high-dimensional,heterogeneous,and mixed features.The rich structural information contained in the extremely inhomogeneous regions is the difficulty in SAR image segmentation and features designed artificially can hardly reflect the essential features of SAR images in the traditional SAR image segmentation method.In this regard,the machine learning model can be used to adptavily characterize the extremely heterogeneous regions of the SAR image.However,the size and quality of the sample set input into the model greatly affect the quality of the extracted features.Therefore,under the guidance of the hierarchical visual semantic model of SAR images,the establishment of sample sets of SAR's extremely inhomogeneous regions is studied.The main tasks are as follows:(1)There is a high degree of overlap between sample blocks collected by every sliding window,and there is a large amount of redundant information.Therefore,an image block selection strategy based on sketch structure clustering is proposed to cluster sample sets collected from every sliding window,and then a sample block with a rich structure is selected from the clustering results to form a sample set to achieve the purpose of removing redundant samples.The distance measurement method based on the sketch structure is defined for the complex ground structure of the extremely inhomogeneous regions of SAR image.Unlike the k-means clustering method,setting the number of clusters artificially is required.Based on the inherent spatial location feature of images,the number of clusters is determined adaptively in our method.On this basis,the results of clustering is compared,based on the distance metrics based on the Euclidean distance and the distance metrics based on the sketch structure.The experimental results show that the distance metrics proposed in this paper can obtain more accurate clustering results.(2)The extremely heterogeneous regions of the SAR image contains complicated ground structure,which are different in shapes and sizes.Beginning with capturing the shape and size of the ground structure,an image block selection strategy based on super-pixel segmentation and sketch line is proposed to construct sample sets for the extremely heterogeneous regions of the SAR image.Because the feature structure and the shadow of SAR images usually appear together,firstly super-pixel segmentation on the SAR image is performed to find the shadow regions,and then the image block containing the ground structure is extracted to form a sample set under the guidance of the sketch line and the shadow.Compared with the method in(1),the number of image blocks extracted by this method is smaller,thus feature learning time is shortened.(3)The semantic information of the sketch line in the hierarchical visual semantic model of the SAR image is not used in the above two schemes,but the direct ground information can be given by the semantic information of the sketch line.The sketch lines aggregated bilaterally generally appear in the extremely inhomogeneous regions of SAR image.Based on this,an image block selection strategy based on the bilateral clustering characteristics of the sketch line is proposed.The sample blocks containing the ground structure of SAR image is fully extracted by extending the sample blocks,thus constituting a sample set for feature learning.The SAR image segmentation is performed after each of the above work,and according to the different characteristics of each pixel subspace in the SAR image,different methods are used to do segmentation,and then the segmentation results of all the pixel subspaces are integrated together to get the final segmentation result of the SAR image.