A Lightweight and Efficient Network for Maritime Object Detection

Maritime object detection is a crucial task in the environmental perception of unmanned surface vehicles. It faces challenges such as complex backgrounds, varying object scales, and numerous small objects. To address these issues, we propose a lightweight network based on YOLOv7-tiny. Our proposed method introduces the P2 detection head with high resolution to improve small object detection, the parameter-free attention module SimAM for better object feature extraction in complex maritime backgrounds, and the CARAFE module to reduce the loss of feature information during upsampling. We conducted experiments on a self-made maritime object detection dataset, and our proposed model outperforms the original YOLOv7-tiny by 2.6% mAP@.5:.95, achieving a real-time inference speed of 101 FPS. Furthermore, our proposed model performs better than other lightweight models.