Contrastive Domain Adaptation-Based Sparse SAR Target Classification under Few-Shot Cases

Due to the imaging mechanism of synthetic aperture radar (SAR), it is difficult and costly acquire abundant labeled SAR images. Moreover, a typical matched filtering (MF) based image faces problems serious noise, sidelobes, clutters, which will bring down accuracy target classification. Different from MF-based result, sparse shows better quality with less noise higher signal-to-noise ratio (SNR). Therefore, theoretically using for classification achieve performance. In this paper, novel contrastive domain adaptation (CDA) method proposed solve problem insufficient samples. method, we firstly construct dataset by complex iterative soft thresholding (BiIST) algorithm. Then, simulated real datasets are simultaneously sent into an unsupervised framework reduce distribution difference obtain reconstructed images subsequent Finally, manually fed shallow convolutional neural network (CNN) along small number Since current definition samples still vague inconsistent, paper defines few-shot as than 20 per class. Experimental results on MSTAR under standard operating conditions (SOC) extended (EOC) show that makes up information limited data. Compared other deep learning methods samples, our able especially few shots.