SAR image target recognition via deep Bayesian generative network

In this letter, a novel deep-leaming-based feature selection method based on Poisson Gamma Belief Network (PGBN), is proposed to extract multi-layer feature from SAR images data. As a deep Bayesian generative network, PGBN has the ability to extract a multilayer structured representation from the complex SAR images owing to the existence of Poisson likelihood and multilayer gamma hidden variables, at the same time the PGBN can be viewed as a deep non-negative matrix factorization model. Note that the PGBN model is an unsupervised deep generative network and it fails to make full use of the label information in training stage. Therefore, the NBPGBN model is further proposed to obtain a higher recognition performance and training efficiency based on Naïve Bayes rule. The experimental results on MSTAR dataset show that the feature extracted by this new approach has better structured information and perform better classification result compared with some related algorithms.