Self-attention mechanism-based SAR for YOLO-v3 maritime ships image target detection

Abstract

In recent years, China's maritime construction has been gradually strengthened, and the security of our territorial waters has become a top priority. In this paper, we propose a self-attentive mechanism-based target detection model for YOLO-v3SAR images, and through experiments, we add a self-attentive mechanism before and after the feature fusion part for target detection, and compare the accuracy, we conclude that adding a self-attentive mechanism before each predicted feature layer can effectively improve the detection accuracy. After adding the self-attention mechanism, the detection accuracy of SSDD dataset increases by 10%, Increased from 84.7 to 94.3%, and that of Ship-dataset dataset increases by 9%, from 79% to 88%. The experiments prove that the improved algorithm model is adapted to SAR image target detection and reaches the advanced level, which provides a new idea for SAR image target detection of maritime ships.