Small Ship Detection of SAR Images Based on Optimized Feature Pyramid and Sample Augmentation

Synthetic aperture radar images have become the latest high-resolution imaging equipment, which can monitor the Earth 24 h a day. More and more deep-learning technologies are applied to ship target detection; however, in complex environments, due to the small target of the ship, problems, such as false detection and miss detection, often occur. For this reason, SSPNet is proposed with several small-target-augmentation strategies to complete the detection of small ships on the sea. This network is an improvement of FPN. The model uses a context attention module (CAM), scale enhancement module (SEM), and scale selection module (SSM). CAM introduces the attention heat map, SEM uses the residual module to make the network pay more attention to specific scale targets, and SSM introduces deep semantic features into shallow features. A weighted negative sampling strategy is proposed to enable the network to select more representative samples. These modules make the network more suitable for small-target detection. The results on the SSDD dataset show that the model is superior to the existing object detection network, and the average precision (AP(50)) reaches 91.57%.

Automated summary

Synthetic aperture radar images have become the latest high-resolution imaging equipment for monitoring the Earth 24/7. The proposed SSPNet utilizes small-target-augmentation strategies and modules such as CAM, SEM, and SSM to improve ship detection in complex environments. The model achieves a superior performance with an average precision (AP(50)) of 91.57% on the SSDD dataset.