WSA-YOLOv5s: improved YOLOv5s based on window self-attention module for ship detection

Ship detection from visual image (SDVI) plays a significant role in terminal management, cross-border ship detection and marine target tracking. Compared to synthetic aperture radar (SDSAR) based ship detection, SDVI has superior performance on accuracy and speed. WSA-YOLOv5s, a new algorithm with a Window Self-Attention (WSA) module is proposed in this paper to substantially improve the ability of detecting small targets while enhancing the ability of large target recognition marginally. The following fine tunings are made based on the original YOLOv5s network. (1) Perform target data enhancement for the dataset’s imbalanced classification by using the improved mix-up algorithm and focal loss; (2) Replace CSP structure in the original backbone with the WSA module, and add one more interaction layer than the original to increase the exchange of high-level semantic information; (3) Add CBAM module to reduce the influence of background interference factors by using spatial attention and channel attention. Experiments are carried out on the Singapore Maritime Dataset (SMD). The results show that our model is greatly improved in all aspects. And the mAP increases from 73 to 90.5%, while keeping the number of model parameters small.