Research On Imbalanced Data Issue In SAR Target Discrimination

Synthetic Aperture Radar(SAR)is a kind of microwave detection sensor,which is not affected by various external environment such as weather and illumination,and it can detect all interested areas in all weather and all day.Since the SAR is successfully developed,the imaging technology of SAR has been developing rapidly,and the resolution of SAR image is getting higher.With the emergence of various SAR systems,SAR has been widely used in various fields.Therefore,SAR target discrimination has also become the research focus of many researchers.However,in practical applications,there are fewer target samples available than the clutter samples,which leads to the imbalanced issue of the training data in the SAR target discrimination.When the training target and clutter data are imbalanced,the target detection rate of the traditional SAR target discrimination algorithm decreases,which leads to the performance of discrimination algorithm decline.The performance of SAR target discrimination algorithm for imbalanced data needs to be improved.To solve the above problems and improve the performance of SAR target discrimination algorithm for imbalanced data,this thesis studies the SAR target discrimination algorithm based on the traditional classification methods of imbalanced data,the SAR target discrimination algorithm based on the cost-sensitive dictionary learning and the SAR target discrimination algorithm based on ensemble learning.The paper is arranged as follows:1.The first part introduces the research background and significance of this thesis,the status of SAR target discrimination research and the status of imbalanced data classification research and the organization of this thesis.2.The second part studies the SAR target discrimination algorithm based on the traditional classification methods of imbalanced data.First,we briefly introduce the process of using the detecion based on two-parameter constant false alarm rate(CFAR)and segmentation algorithm based on super-pixel to obtain the image chips.Then,the extraction method of discrimination feature based on bag of word model is briefly described and the experiment is carried out on the mini SAR data set.Finally,some traditional methods for the classification of imbalanced data set are studied,and the comparison experiments are carried out on the mini SAR data set.3.The third part studies the SAR target discrimination algorithm based on the cost-sensitive dictionary learning.First,we introduce the basic theory of sparse representation classification and K-SVD dictionary learning algorithm.Then the framework and process of the SAR target discrimination algorithm based on cost-sensitive dictionary learning is introduced.The cost information is considered in the sparse coding phase,and enforce the cost-sensitive constraints in the process of the whole dictionary learning.Finally,the comparison experiments are carried out on the mini SAR data set.The results verify the cost information can improve the performance of SAR target discrimination algorithm for imbalanced data.4.The fourth part studies the SAR target discrimination algorithm based on ensemble learning.First,we introduces the basic theory of ensemble learning.Then,we describe the SAR target discrimination algorithm based on three different ensemble methods of dictionaries,and the comparison experiment is carried out on the mini SAR data set.Finally,the framework and process of SAR target discrimination algorithm based on ensemble multiple classifiers are introduced.The base classifiers of the algorithm are respectively sparse representation classifier based on cost-sensitive dictionary leaning(CSDL-SRC)and support vetor machine(SVM),and the comparison experiments are carried out on the mini SAR data set.The experimental results verify the ensemble learning can further imporve the performance of the SAR target discrimination algorithm for imbalanced data.