Unsupervised Domain Adaptation for SAR Target Detection

Recent years have witnessed great progress in synthetic aperture radar (SAR) target detection methods based on deep learning. However, these generally assume the training data and test obey same distribution, which does not always hold when parameters, imaging algorithm, viewpoints, scenes, etc., change practice. When such a distribution mismatch occurs, it will cause significant performance drop. Domain adaptation provide an effective way to address this problem by transferring knowledge from source domain (training data) (test data). In article, we proposed unsupervised faster R-CNN SAR framework adaptation, can improve unlabeled borrowing of labeled domain. Our approach is composed following three stages: pixel-domain (PDA), multilevel feature (MFDA), iterative pseudolabeling (IPL). By generating transition using generative adversarial networks, PDA stage reduce appearance differences images. At MFDA stage, detector only learn domain-invariant global features instance-level regional via learning common space but also reweight low-level according their relative importance IPL design pseudo labeling strategy that select pseudo-labels instance level image encourage more discriminative directly. We evaluate our method miniSAR FARADSAR datasets. The experimental results demonstrate effectiveness approach.