Classification for Synthetic Aperture Radar Ship Image of the Ship Targets under the Few-shot Condition

Abstract

Deep learning algorithms have shown outstanding results for the categorization of synthetic aperture radar (SAR) images. But because there is a conflict between the deep learning techniques' broad parameter space and the scanty labeled samples of ship targets, most deep learning models are unable to provide satisfying outcomes under the few-shot scenario. Classification for SAR ship images often faces the few-shot condition because they are sensitive to system parameters, so that the amount of effective SAR images is very low. To this end, this article proposes a method based on graph neural network (GNN). An end-to-end classification model was trained by this method. Moreover, our method extracts multi-scale features of the targets to make classification effect better. The dropout layers are added into the graph convolution layer to reduce the overfitting in few-shot classification. Finally, our model is applied to the mixed dataset of the simulated and real data to verify effectiveness of our method. The experimental findings presented in this research show that our method outperforms other cutting-edge approaches in terms of categorization rate.