Abstract: Compared with optical images with a large number of labeled data， synthetic aperture radar （SAR） images lacked sufficient labeled samples， which limited the performance of supervised learning SAR target recognition algorithm. However， the unsupervised recognition methods were difficult to meet the practical needs. This paper proposed a semi-supervised generative adversarial network based on self-attention feature fusion，alleviating the challenge of annotated data lacking. Firstly， the self-attention layer was introduced in the construction of generator and discriminator to overcome the problem that convolution did not have long range extraction capability. Secondly， the discriminator utilized multi-level feature fusion to capture the key information of SAR image. Finally， spectral normalization was applied in the training process to improve the convergence stability. In order to verify the effectiveness of the proposed method， experiments were carried out on Moving and Stationary Target Acquisition and Recognition （MSTAR） data sets. Experiment show that the proposed method can learn valuable information from unlabeled samples， effectively solve the problem of insufficient labeling.