 We proposed a novel approach for solving the problem of limited labeled SAR target data. Firstly, we trained a deep convolutional autoencoder, DCAE, using a large number of unlabeled CRC images. This DCAE was then used as a pre-trained model to transfer knowledge to SAR target classification tasks. Additionally, we introduced a feedback bypass with reconstruction loss to further improve the classification accuracy. Experimental results showed that our approach outperformed other methods when dealing with limited labeled SAR target data. This article was authored by Xiong Ling Huang, Xiong Xiu Pan, and Bin Lei.