Target recognition for SAR images based on heterogeneous CNN ensemble

Limited SAR training data could incur overfitting in multilayer convolutional neural network (CNN), which restricts the recognition performance severely. In this paper, we proposed a new SAR target recognition method based on heterogeneous CNN ensemble. The proposed network is constructed with the noncomplete connection scheme and multiple filters stack, which is able to reduce the number of free parameters and improve the training efficiency. Based on ensemble learning theory, heterogeneous network structures are effectively combined, the generalization ability and recognition performance of the network are improved, and the amount of training samples is also further reduced. Experimental results have shown the superiority of the proposed network based on the MSTAR dataset.