Multi-Scale Dense Networks for Ship Classification Using Dual-Polarization SAR Images

Abstract

As one of crucial remote sensing applications, ship classification using synthetic aperture radar (SAR) images has increasingly been studied in modern maritime surveillance. Nowadays, the prevailing classification paradigm for SAR ship targets is to utilize the deep network models, which presents superior performance over the traditional handcrafted feature driven methods. Of which the SAR ship classification method using densely connected convolutional neural networks (CNNs) is among the state-of-the-art. However, the general CNNs cannot fully explore the SAR ship feature representations, which limits its potentials for better classification performance. In this paper, we propose a novel multi-scale framework for the CNNs to further improve the ship classification performance with dual-polarization SAR images. Particularly, the convolutional feature maps from different spatial scales are fused to acquire multi-scale global representations of the dual-polarization SAR images, which are finally integrated by the group bilinear pooling operation in the classification layer and will further be processed by multiple classifiers for better network training. Extensive experiments have proved that the proposed method can improve the robustness and classification performance against the state-of-the-art algorithms on the OpenSARShip datasets.