SAR Object Classification with a Multi-Scale Convolutional Auto-Encoder

摘要

Despite of the significant success obtained by the deep networks, insufficient labelled training has often been the major problem while applying the deep learning models in SAR object classification tasks. In this paper, an unsupervised deep learning model that is implemented in the encoding-decoding architecture is proposed. The proposed deep network learns feature maps at different scales and combined them together to generate feature vectors for object classification. Besides, the reconstruction loss is improved by computing the mean square error between the reconstructed images and the data processed by an improved Lee Sigma (ILS) filter so that the background clutter in the target patches can be suppressed. The open published MSTAR dataset is utilized for performance evaluation. Both the validation results and comparison experiments demonstrates that the proposed model can adaptively learn discriminatory features from raw SAR data.