Research On General Gragh Embedding Manifold Learning For SAR Target Recognition Methods

Synthetic Aperture Radar(SAR)can obtain high-dimensional and high-resolution images of the Earth's surface all-time and all-weather,and has wide applications in the fields of agriculture,geological exploration,disaster assessment,and military reconnaissance,etc.The automatic target recognition(ATR)technology of SAR can automatically,accurately and efficiently classify and recognize the targets in SAR images,which has great research and application value.It is currently the front-end hotspot of international SAR technology.Feature extraction is a key step for SAR ATR.Based on the distribution characteristics of SAR image datasets and the theory of manifold learning,this dissertation develops theoretical analysis,method research and simulation verification of feature extraction methods for SAR targets.The main innovations are as follows:(1)A type of MSDE method was proposed.By introducing the marginal sample discriminant coificient,the marginal samples were better clustered,and the misclassification caused by marginal samples in the sample aliasing region was solved.Thus,the recognition rate of the proposed method was improved.(2)A type of sample-entropy based distance discriminant analysis method was proposed,which not only enabled neighbor between-class samples to be separated and non-neighbor within-class samples to be clustered,but also can incorporate sample contribution when feature extraction was performed.The proposed method achieved better recognition performance in a lower dimensional feature space.(3)A type of local sample directional discriminant projection method was proposed,which can build a neighbor clustering center based on the entropy of neighbor information of each sample and indicate the clustering direction.Thus,it achieved better clustering in the low dimensional space,and improved the recognition rate.(4)The framework of the generalized graph embedding manifold learning method was proposed,which revealed the common physical nature of the above three manifold learning methods and the existing manifold algorithms,and showed the transferability of the mathematical representation,meanwile achievd a unified mathematical representation of the feature extraction strategy.The framework layed a foundation for the application of manifold learning method applying in SAR ATR system.The above methods have been verified and simulated by experiments on MSTAR database.The results showed that the SAR feature extraction algorithms proposed in this thesis can solve small samples problem and nonlinear problem.They can be applied to different types of features in vector space,kernel space and tensor space,and can effectively improve the recognition accuracy and recognition stability of SAR image recognition.This work improved the theory of manifold learning and layed a theoretical foundation for its efficient implementation in engineering applications.