Although deep neural network has achieved great success in Automatic Target Recognition with Synthetic Aperture Radar (SAR-ATR), there are still two major problems in the existing works. Firstly, most networks belong to supervised learning, which has a great requirement on the number of target labels. Secondly, the phase information of complex-valued images is generally discarded regardless of their significance, and only the amplitude information is used. In order to solve these problems, this paper proposes a model for vehicle SAR image interpretation based on complex-valued convolutional auto-encoder (CV-CAE). Several advanced network structures have been extended to complex-valued domain and utilized in the proposed model to improve its performance, including the Inception structure, the residual block and the attention mechanism. The proposed network not only learns features in an unsupervised manner, but also makes full use of the phase information in SAR images. Finally, the KNN classifier is adopted to categorize the vehicles based on the features learned by the proposed CV-CAE. Experimental results with the MSTAR dataset indicate that the proposed method achieved a comparable performance to the state-of-the-art results.