A Hierarchical Feature Fusion and Attention Network for Automatic Ship Detection From SAR Images

Abstract

Automatic ship target detection technique is a typical and meaningful application for synthetic aperture radar (SAR) image interpretation. Nevertheless, the detection of ship targets within SAR imagery is encumbered by several detracting elements, including obscured outlines, varying dimensions, and elaborate backgrounds, which collectively render the identification process challenging. Existing methodologies for discerning ship targets prove inadequate in effectively navigating these complications. Therefore, we propose a new deep neural network to automatically detect ship target from SAR images, which is named as hierarchical feature fusion and attention network (HFFANet). HFFANet is based on CSPDarknet, the backbone network of YOLOX, and adaptive feature fusion and attention (AFFA) module is innovated to enhance feature extraction. In AFFA, adaptive multilevel feature fusion module is proposed to achieve effective multilevel feature adaptive fusion to better extract target contours and suppress background clutter to reduce false alarms, and enhanced residual coordinate attention module is also proposed to enhance spatial location information and embed it into channel features in the channel layer. The experiments on SAR ship dataset have been conducted, and the mean average precision of HFFANet is 98.53%. Compared with the classical networks, the experimental results show that our model not only achieves the optimal balance in precision and recall, but also achieves the optimal calculation cost.

Keywords

• Attention mechanism
• deep learning
• multilevel feature fusion
• ship target detection
• synthetic aperture radar (SAR) image interpretation