Accurate classification of missile by the missile image (or in flight state) taken through the satellite equipment, which achieve the timely and effective defense, is one of the hot spot in the military field at home and abroad. Because the missile in the war state has masked color, and the missile shape differences are not significant, it is difficult to classify the missile type based on the low level features. Aiming at these problems, a new algorithm was presented based on Sparse Auto-Encoder (SAE) combining the high level visual feature and low level feature extraction. In order to improve classification accuracy, transfer learning was introduced, with the help of the STL-10 sample database local features, the global features of small sample missile target image can be extracted through the local features by the convolution neural network (CNN) of pooling layer, and then transmitted into the Softmax regression model to realize classification of missiles. Experiments show that compared with the traditional low level vision features and SAE high level vision feature classification algorithm, the SAE fusion feature classification algorithm has higher accuracy and robustness. In addition, in order to avoid classification performance reduce even failure under the lack of training for new type missile target object, the new algorithm induces transfer learning to extract local feature, experimental result proves the feasibility and accuracy of the algorithm.