Ship Detection in Inland Rivers Based on Multi-Head Self-Attention
Nanjing Yu, Xiaobiao Fan, Tianmin Deng, Guotao Mao · 2022 7th International Conference on Signal and Image Processing (ICSIP) · 2022
Ship object detection is becoming a key technology for unmanned ships and plays a crucial role in ship safety. Aiming at the characteristics of a large difference in scale between classes and many small objects in inland rivers and ports, this paper proposes MHSA-YOLO (Multi-Head Self-Attention-You Only Look Once) based on Multi-Head Self-Attention (MHSA) and You Only Look Once (YOLO) network to implement the ship object detection in inland rivers. More specifically, this proposed method introduces MHSA in the feature extraction process to weaken the interference of complex background information and strengthen the feature information of the ship objects. Further, a simplified two-way feature pyramid is developed in the feature fusion process so as to strengthen the fusion of feature information and the representation ability. The experimental results on the Seaships dataset show that the proposed MHSA-YOLO method achieves 97.59% [email protected], and is more effective compared with the state-of-the-art (SOTA) methods.