Abstract: Synthetic Aperture Radar (SAR) can obtain SAR images of the target from a number of different azi-muths when observing a ground target, but the shapes of the target in these images are different. In view of the fact that SAR image is extremely sensitive to the observation azimuth and small scale of SAR im-age dataset, this paper designs a Convolutional Neural Network (CNN) for multi-aspect SAR images target recognition. Three SAR images of the same target are regarded as a pseudo-color image inputted to the network, which making full use of the acquisition characteristics of SAR image data. Instead of flat-tening, we use pool layer to reduce the number of parameters of network at the same time. The experi-mental results show that this convolutional network architecture has high recognition precision on small scale of SAR dataset, and has excellent recognition performance for different types of targets in the same category.