Cross Sensor Transfer Learning for Unsupervised SAR Target Detection

Abstract

Synthetic Aperture Radar (SAR) target detection based on deep learning typically assume the training and test data drawn from the same distribution. However, when the radar sensors change, this assumption does not always hold and will cause to a performance degradation when encountering such distribution mismatch. Transfer learning shares a great idea to solve this issue by transferring knowledge from the labeled data of source sensor to the unlabeled data of target sensor. We proposed a cross sensor transfer learning method based on domain adaptation, which can improve the cross sensor robustness of SAR target detection with unlabeled target sensor (domain) data. Our method includes three stages: pixel domain diverse (PDD), multilevel feature alignment (MFA) and iterative self-training (IST). The PDD stage can alleviate the few training data of SAR target detection and enhance the generalization ability by generating transition domain. MFA can learn domain invariant features in a multi-layer feature space by using the idea of adversarial learning. To learn more discriminative features, we further proposed the IST that can generate and select high-quality pseudo labels for the SAR images of target sensor. The experimental results on miniSAR and FARADSAR datasets demonstrate the effectiveness of the proposed approach.