Underwater Sonar Image Targets Detection Based on Improved RT-DETR

Abstract

Underwater sonar imagery is characterized by small target sizes and low resolution, which can result in detection failures or false positives. To counteract these challenges, we introduce the underwater sonar detection transformer (US-DETR), an underwater sonar object detection model derived from the real-time detection transformer (RT-DETR) framework, incorporating attention-based feature fusion. US-DETR includes a novel enhanced feature interaction (EFI) module, which enhances the feature extraction network’s ability to perceive global information of the detected target. In addition, we propose a novel nonlocal attention feature fusion (NAFF) module to heighten the network’s sensitivity to the spatial relationships between feature channels across different scales, thereby enhancing its channel position and global information awareness. Experiments are conducted on a benchmark underwater sonar image dataset. Experimental results show that compared with RT-DETR, US-DETR achieves a 2.2% higher mean average precision (mAP) and a 2.1% higher $F1$ score compared with RT-DETR. The model also strikes an effective balance between detection speed and accuracy, achieving real-time performance of 126 FPS, which can meet the real-time requirements in industrial production.