Research on Detection Algorithms Based on RT-DETR Networks for Ship Objects in SAR Images

In Synthetic Aperture Radar (SAR) ship target detection, the challenges of misdetection and low detection accuracy for small ship targets, as well as the difficulty in distinguishing small ship targets from similar background objects, have prompted the proposal of a SAR ship detection network called CG-DETR, built upon the real-time detection transformer (RTDETR). This network introduces a new backbone network: BCNet, which employs a deep convolutional structure to facilitate the extraction of more detailed features, thereby enhancing the detection of small ship targets. Additionally, a content-guided module is designed and integrated into a cross-scale fusion module (CCFM), which achieves deep feature fusion between different feature maps through context guidance and weighting operations. This approach enhances the capture of critical information while mitigating background interference. The model is validated on the public dataset HRSID, and experimental results demonstrate that the CG-DETR network improves overall detection accuracy by 2.7% while maintaining a similar number of parameters and computational load, Notably, it achieves a 3.8% improvement in the detection accuracy of small ship targets.