Few-Shot SAR ATR Based on Conv-BiLSTM Prototypical Networks

In recent studies, deep neural network has been successfully applied to synthetic aperture radar (SAR) automatic target recognition (ATR). However, these algorithms require hundreds of training samples of each class of targets that need to be recognized. In order to recognize the target with only a few training samples, this paper proposes a new few-shot SAR ATR method based on Conv-BiLSTM Prototypical Networks (CBLPN). First, a Conv-BiLSTM network is trained to map SAR images into a new feature space where it is easier for classification. Then, a classifier based on Euclidean distance is utilized to obtain the recognition results. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) benchmark dataset illustrate that the proposed method achieves an accuracy of over 90% on the classification of three classes of targets with only 5 training samples of each class.