TSDet: End-to-End Method with Transformer for SAR Ship Detection
Yanyu Chen, Zhihao Xia, Jian Liu, Chenwei Wu · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
Synthetic Aperture Radar (SAR) ship object detection has important applications in maritime monitoring. Currently, CNN-based detectors are hardly perfect for multi-scale SAR ship detection by scattered noise and complex land and sea background interference. Therefore, we propose a new end-to-end method with Transformer for SAR ship detection (named TSDet). TSDet adopts an anchor-free design that can directly predict object bounding boxes and classes of multi-scale ships. Specifically, we propose Perceptual Enhancement Transformer (PET) at first, as the network structure of TSDet, which consists of the Feature Perception Enhancement (FPE) module based on self-attention and the Task Perception Enhancement (TPE) module based on cross-attention. PET can effectively suppress background noise in SAR images and enhance the salient features of ships. Second, to solve the sparse features of SAR ships, we propose a new effective sparse attention alternative to the original attention module in PET networks. It can quickly focus on important information in global features and accelerate network convergence. Finally, on the SAR-Ship dataset, our TSDet achieves the best 93.77 % AP50, outperforming competitive benchmarks.