SAR Image Segmentation Based On Semantic Space And Deconvolutional Networks Learning Model

SAR image segmentation is an important job of SAR image processing.The result of the segmentation will affect the object detection,recognition and following processing directly.The traditional segmentation of SAR image mainly includes the method based on spatial domain and the method based on feature domain.But because SAR image contains much speckle noise;and SAR image has the characters of heterogeneity in high-dimension and contamination.The traditional feature extraction often spends much time and labor.Our team proposed the hierarchical semantic space on the basic of the semantic information of SAR image.Under the guidance of hierarchical semantic space,this thesis uses the popular Deconvolutional Network to learn the structural feature of the image.Then we use the Primal Sketch to extract sparse structural information in image.At last,we combine the semantic information and the method proposed in this thesis to realize the segmentation of SAR image.The main work of this thesis is follows:(1)We select the curvelet filter of different direction to initialize the filter group in each layer of deconvolutional network,getting the curvelet deconvolutional network.Because curvelet filter contains the information of direction,scale and position,it can match the structural information in image precisely.Compared with the method of random initializing filter and Gaussian initializing filter,using the curvelet filter to initialize the filter group in deconvolutional network can help the network learn better image feature.At the same time,the learning speed will improve quickly.(2)The thesis proposes the segmentation method of gathering pixel space based on Deconvolutional Network and the direction constraint of sketch line.Firstly,under the guidance of the hierarchical semantic space,we divide the entire SAR image into gathering pixel space,structural pixel space and homogeneous pixel space.For each highly unhomogeneous area in gathering pixel space,we respectively train a curvelet deconvolutional network,then,we use the network to learn the feature of different area,getting the filter set of each area.Secondly,we use the Primal Sketch model handle the filter set,getting the sketch block corresponding to the filter in set.Thirdly,in order to represent the directional feature of filter,we design the directional feature vector on the basis of the sketch line in sketch block.Fourthly,we design the directional cluster method based on directional feature vector and xor operation to cluster the filter set in the guide of direction.At last,we combine the directional clustering result of filter set in different area,regarding it as the codebook,then,we design the method of directional codebook mapping to get the structural feature of each area,realizing the segmentation of gathering pixel space.(3)The thesis proposes the segmentation method based on visual semantic space and Deconvolutional Network learning model.For the gathering pixel space,we use the segmentation method of gathering pixel space based on deconvolutional network and the direction constraint of segment.For the structural pixel space,we use the visual and semantic rule to segment the line object and use the sketch gathering rule to segment the independent object.For the homogeneous pixel space,we use the polynomial model based on self-adapted neighborhood to segment it.At last,we combine the segmentation results of all the space,getting the final SAR segmentation result.