Research On High Resolution PolSAR Images Classification Based On Mid-level Semantic Features Model

Polarized Synthetic Aperture Radar(PolSAR)can acquire more abundant surface information all day and multi-polarization by utilizing the penetration of electromagnetic wave and the combination of different polarization mode.Therefore,it is widely used in target detection,land-use classification,surface parameter inversion and other fields.With the increasing resolution of the image acquired by the PolSAR system,the traditional pixel-level interpretation methods are restricted to some extent.Therefore,how to interpret high-resolution PolSAR images quickly and accurately becomes the focus of the research.In this thesis,a method of PolSAR image classification based on mid-level semantic feature model is studied in order to solve the image interpretation problems based on low pixel-level feature.The object information of the high-resolution PolSAR image is more complicated and rich,and the processing time based on pixels is longer.Therefore,the segmentation operation is performed to obtain the object with good edge as the processing unit to improve the interpretation speed effectively.Computers usually acquire low-level features for interpretation.The low-level features of PolSAR images are pixel-based polarization features obtained by target decomposition.High-level features are partial or global features based on images that combine object information,semantic context information,scale space information,etc.,and the features are closer to people's understanding.The introduction of the mid-level features model can establish a bridge between the low-level features and the high-level features to make the computer's interpretation process closer to the process of people understanding the image.Both the BOW model and the pLSA model are mid-level feature models,which transform the abstract low-level polarization features into the concrete “words” and “themes” mid-level features to understand image betterly.This paper first introduces the domestic and foreign research status of PolSAR image feature extraction,mid-level semantic feature modeling and classification.Focused on the special imaging mechanism of PolSAR image and the realization of land-use classification,this paper introduces the segmentation method based on region growing and merging to obtain the segmentation region as the smallest unit for mid-level feature representation.Secondly,the BOW model is used to extract the mid-level semantic features and implement classification in each acquired segmentation region.In this paper,two methods are used to construct visual words.One is to form a visual dictionary based on K-means method,and the other is to construct multi-scale visual dictionary based on multi-scale context information and prior knowledge information.The image area is characterized by visual words and combined with the SVM classifier to achieve classification.Then the pLSA model is introduced to set the potential relationship between words and documents,and the mapping of visual words to images is transformed into the mapping of implicit semantic topics to images,which effectively reduces the feature dimensions and improves the interpretation accuracy.Finally,two groups of PolSAR data are used for land-use classification experiments,and a group of PolSAR data is used for scene classification experiments to verify and analyze.The experimental results show that the proposed method can effectively improve the classification speed of high-resolution PolSAR images on the premise of ensuring classification accuracy,and verify the rationality and practical applicability of the method.