PolSAR Image Segmentation Based On Feature Learning And Low-rank Decomposition

The characteristic of Polarimetric Synthetic Aperture Radar(Pol SAR)is that it did not limited by the climate and environment conditions, which can observe all-day, and plays an important role in the field of military and civilian nowadays. As Pol SAR image segmentation is an important part of Pol SAR image processing and interpretation, the segmentation quality can directly affects the accuracy of subsequent Pol SAR target detection and recognition. Without consideration of the characteristics of structure and spatial relationship between pixels, and based on the characteristics of the pixel, the traditional Pol SAR image segmentation methods the did not act well in the consistency of the area segmentation. Based on the regional map, this paper using deconvolution network and low-rank decomposition extract the relationship between pixel, the main research results are as follows,1) Put forward a Pol SAR image segmentation method based on the low-rank decomposition and histogram statistics. According to SAR sketch model to extract power figure and area figure of Pol SAR images, and mapping the Pol SAR image as concentrated areas, uniform areas and structure areas based on the regional map, through the segmentation operation of the regional map of the concentrated areas, uniform areas and structure areas respectively to get the final segmentation result. Aim at the characteristics of a strong clustering and structural relationships in the concentrated areas, adopts the low-rank decomposition model to deal with the concentrated areas which is not connect on the space and get the histogram statistics of the low rank part of low-rank decomposition, use the Bhattacharyya Distance to compute the distance between different concentrated areas and construct the similarity matrix, finally based on the spectral clustering method and graph cut using similarity matrix to combine the concentrated areas to get the final segmentation result of the concentrated areas. As the uniform area does not have obvious structural relationships, using H/alpha/A- wishart segmentation method to segment the uniform area. For structure area using the watershed segmentation and Wishart distance super pixels merging method to segment, and combine the segmentation result with the segmentation result of concentrated areas and uniform areas to get the final segmentation result.2) Put forward a Pol SAR image segmentation method based on the deconvolution network and sparse classification. On the basis of the Pol SAR regional figure, extract the Pol SAR coherent matrix diagonal three elements of amplitude value and upper triangular matrix phase value of the three elements, for the every disconnect concentrated area, using three-channel amplitude value and three-channel phase value training a deconvolution network with 4 layer respectively, to extract the last layer of all filters as a group of filter, to merge the range filter and phase filter for each concentrated area, to construct the dictionary using the filter of every concentrated areas, project the filter of each concentrated area for the dictionary respectively, and calculating the average projection as the feature vector of the concentrated areas, calculate the cosine distance between every two concentrated areas and construct the similarity matrix, finally based on the spectral clustering method and graph cut using similarity matrix to get the final segmentation result. This method is mainly focus on the concentrated area, using H/alpha/A-wishart segmentation method to segment the uniform area. For structure area using the watershed segmentation and Wishart distance super pixels merging method to segment, and combine the segmentation result with the segmentation result of concentrated areas and uniform areas to get the final segmentation result.