A Fully End-to-End Query-Based Detector with Transformers for Multiscale Ship Detection in SAR Images

Recently, deep learning-based ship detection methods have shown higher accuracy, stronger robustness, and better generalization ability than traditional methods in complex maritime scenes. However, these deep learning-based methods rely heavily on the anchor mechanism and non-maximum suppression (NMS) operation, which increases the complexity of the detection process. Furthermore, some existing studies place too much emphasis on detecting small ships and neglect the same importance of detecting medium and large ships. In this paper, we introduce a fully end-to-end query-based detector with Transformers for multiscale ship detection in synthetic aperture radar (SAR) images. Our model, based on the DEtection TRansformer (DETR) framework, utilizes learnable queries and bipartite graph matching instead of hand-crafted anchors for set-based direct box prediction, resulting in end-to-end optimization. In addition, considering the multiscale characteristics of SAR ships, a balanced deformable attention module (BDAM) is proposed to extract contextual semantic information and accelerate DETR training convergence. Through extensive experiments conducted on the SSDD and HRSID datasets, our method demonstrates a competitive performance, surpassing other state-of-the-art approaches.