SA<inline-formula><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula>Net: Ship Augmented Attention Network for Ship Recognition in SAR Images

Abstract

Maritime surveillance is extensively concerned by worldwide authorities, in which ship recognition in synthetic aperture radar (SAR) images is a significant and fundamental component. Though some development has been achieved in the SAR ship recognition task, two areas remain inadequately explored, which are the comprehensive utilization of multiscale features and the deployment of the prior knowledge of the ship shape. In this article, a novel ship augmented attention network (SA$^{2}$Net) for ship recognition is proposed, which comprehensively utilizes the multiscale features and integrates the ship shape prior to the end-to-end network. On one hand, due to the unequal effects of different scales, a scale attention module is proposed to adaptively select and assign weights to desired feature scales while disregarding irrelevant scales. Moreover, a feature weaving module (FWM) is constructed to merge semantic and detailed features produced by the high-to-low backbone, enriching representations across all scales of ship targets. On the other hand, in order to incorporate the priory knowledge of the ship shape into the network, we develop a feature augmentation module (FAM) to further boost the ship recognition accuracy. This module can provide rectangular receptive fields that align with the shape of ships, wherein a limitation encountered with traditional square convolutions. Comprehensive experiments on representative three- and six-category OpenSARShip tasks and seven-category FUSAR-Ship tasks show that our SA$^{2}$Net demonstrates superior performance when compared to the current state-of-the-art methods.

Keywords

• Synthetic aperture radar (SAR)
• ship recognition
• convolutional neural networks (CNNs)
• shape priory knowledge
• feature augmented module
• scale attention module (SAM)