Automatic target recognition is a crucial task for SAR remote sensing. Unlike other methods, the unsupervised representation learning based on deep architecture can obtain robust high-level features directly from raw data. A drawback of most unsupervised representation learning methods in SAR ATR is that they only deal with amplitude images. In addition, many methods utlize a single layer architecture to extract pixel-level/mid-level features which are probably sensitive to condition variation. In this paper, a feature extraction method based on modified stacked convolutional denoising auto-encoder (MSCDAE) for complex SAR images is proposed, where convolutional kernels of MSCDAE are learned by 1-D modified denoising auto-encoders. By stacking the convolutional layers and pooling layers, high-level representation of objects are learned. The features are subsequently sent to a trained SVM for object classification. Experimental results demonstrate that the proposed method can provide a significant improvement in the ATR performance.