SAR Target Discrimination Based On Autoencoders

Synthetic Aperture Radar(SAR)is a high resolution radar that is not affected by environment,such as illumination and climate,and is widely used in military and civilian fields.Among the many applications of SAR,the automatic target recognition(ATR)technology of SAR image has very important military value.The technology can effectively interpret effective information in the target,improve the combat capability and intelligence analysis capability of the army.SAR image target discrimination is an important part of SAR image target recognition technology,and SAR image feature extraction is also a key step of SAR target discrimination.Whether its feature extraction is good or bad will directly affect the accuracy of SAR target discrimination and recognition results.The nonlinear transformation method is an important branch of the feature extraction for this method can excavate the potential information of the data and realize the dimensionality compression of the high dimensional data.The deep learning model,which contains powerful nonlinear mapping relations,can effectively and automatically learn the expression of features in high-dimensional data.It has become a research hotspot in recent years.With the rapid development of SAR imaging technology,the increasing high resolution SAR images can be obtained,and the information in SAR images have become more complicated.Therefore,introducing deep learning into the field of SAR ATR is necessary and urgent in this era of high-speed information and intelligence.Combining deep learning theory,this research studies the SAR target discrimination algorithm based on the autoencoders.The main contents of each section of this dissertation are summarized as follows.1.This thesis introduces the background,significance and development of SAR image target discrimination and target recognition,at the same time,this part summarizes the main work of the thesis.2.The second part in this thesis mainly studies SAR target discrimination based on sparse autoencoders.The feature extraction of SAR is usually the manual extraction of discriminative features based on expert knowledge.This process is time-consuming and laborious,and the acquired features will lose some information of the original data.The sparse autoencoder can automatically learn the features of the data and reduce the human interference.The algorithm of the SAR target discrimination in this part use SAE,which is an unsupervised learning method that can automatically learn features from SAR image.And the features are classified using Softmax classifier.This part verifies the feasibility and effectiveness of the features extracted by SAE.3.For the classification task,the third part studies class encoder with the same structure as autoencoder.Class encoder,whose training objective is to reconstruct a sample from another one of which the labels are identical,aims to minimize the intra-class variations.Class encoder can learn the discriminable features of SAR images and achieves “task oriented feature learning”.This part further studies class-encoding classifier(CEC)which impose the class encoder as a constraint into the Softmax.CEC performs reconstruction tasks and recognition tasks simultaneously in a network model,which make the learning process is in favour of feature extraction as well as classification.4.This section studies SAR target discrimination algorithm based on stacked sparse autoencoder at the beginning,then studies SAR target discrimination algorithm based on stacked sparse autoencoder which is imposed the Fisher criterion,as a constraint into the SSAE,it can describe the relationship between intra-class features and the inter-class features.The Fisher criterion as a constraint on the deep features in the SSAE fine-tuning process can keep the features in small distance within-class and large distance between-class,so that the deep features learned are more conducive to SAR target discrimination.