Semi-Supervised SAR Target Detection with Cross-Domain Transfer Learning Based on YOLOv5

Deep learning-based Synthetic aperture radar (SAR) target detection methods have been a research hot-spot. But they rely on a large number of labels to obtain satisfying results. SAR images are hard to be interpreted, and thus the labels are hard to be collected. In this paper, we develop a semi-supervised method combining with cross-domain training learning based on the YOLOv5 for SAR target detection. In detail, the related cross-domain data from optical images are introduced to assist feature learning of SAR data domain by the way of feature domain adaption constraint. Thus, the abundant related knowledge of optical data are transferred to SAR data. Moreover, we design a decoding module to make use of unlabeled SAR images by data reconstruction, which can exploit much useful information to assist target detection. Results on measured FARADSAR dataset show that our method can achieve promising detection performance under limited labeled data.