The Ship Detection Method of SAR Image Based on Improved YOLOv5
Zhihe Chen, Yuxin Wang · 2023
To solve the issue of ship detection in synthetic aperture radar (SAR) images from mainstream convolutional neural object detection networks due to dense ship distribution, we propose an improved ship detection method for SAR using YOLOv5s. We replace the C3 module with a Swin Transformer in the feature extraction network of YOLOv5s and use a Transformer module to model the global information of the convolutionally extracted object features for detection, enhancing ship detection capability in SAR images. Experiments are conducted on publicly available SAR image ship detection dataset resulting in an average mean accuracy improvement from 94.3% to 96.0%, and network model parameters reduced by 230801, showing the superiority and effectiveness of our proposed method with better detection speed and network complexity than other algorithms.