Polarization Matters: on Bilinear Convolutional Neural Networks for Ship Classification from Synthetic Aperture Radar Images

Ship classification using synthetic aperture radar (SAR) images plays a core role in modern maritime surveillance. Traditional classification methods mainly applied the handcrafted features for ship representation in SAR images, which can hardly deal with the resolution-limited SAR images well. Recently, deep learning methodology opens a new door for effective and efficient SAR ship classification. This paper is dedicated to make a deep exploration on the convolutional neural networks (CNNs) for SAR ship classification. First, a novel two stream CNN framework is proposed to sufficiently utilize the polarization information in SAR images. Second, bilinear pooling is applied to the deep CNN features from two single polarization SAR images, hence more discriminative ship representations are acquired. Third, to mitigate the defect of small-scale SAR ship data set for training the deep CNNs, we propose to finetune the large-scale data pretrained CNN models by task-specific SAR ship data. Extensive experiments demonstrate the effectiveness of the proposed SAR ship classification framework.