Adapting Intra-Class Variations For Sar Image Classification

This paper presents a semi-supervised domain adaptation (SSDA) method for Synthetic Aperture Radar (SAR) image classification. SAR imagery is important in ground activity monitoring, but its wide application is impeded due to a lack of annotations. SSDA methods transfer class-discriminative knowledge from a fully-labeled source dataset to a scarcely-labeled target dataset. However, conventional methods often train models which overfit to labeled target data and fail on unlabeled data. To overcome this, we propose to additionally adapt intra-class variations. Specifically, a conversion network is trained to learn from source data the image feature variations caused by the change of image capturing angle. Then synthetic data, which represent a generalized target domain distribution, are estimated by applying the conversion to labeled target data. Our method improves the accuracy of the state-of-the-art SSDA approach from 64.28% to 80.40% in three-shot cases on the SAR ground vehicle dataset.