DCEA: DETR With Concentrated Deformable Attention for End-to-End Ship Detection in SAR Images

Abstract

Recently, significant advancements have been achieved in optimizing algorithms for synthetic aperture radar (SAR) ship detection. Nevertheless, two challenges still impede further research as follows. 1) Mainstream methods, whether anchor-free or anchor-based, adhere to a dense paradigm, leading to substantial redundancy and limited adaptability. 2) Ship targets in SAR images exhibit large shape variations and scale differences, making it difficult to efficiently extract key features from background clutter. To tackle the aforementioned problems, we propose DETR with Concentrated dEformable Attention (DCEA), a query-based method for end-to-end optimization of the current pipeline. First, for the irregular shapes and sparse distribution of ships, the concentrated deformable attention is introduced to model the spatial positions of targets, simulating their geometric transformations with precision. Second, an attentionwise propagation module is designed to integrate local fine-grained information with global semantic information, improving the detection performance for objects across diverse scales. Finally, due to the lack of information exchange between object queries, a dimensionwise information mixing module is employed to incorporate key information from various dimensions to enhance their representation capability. To validate the superior performance of DCEA, we conduct extensive experiments on multiple public datasets, achieving mean average precision scores of 0.991, 0.929, and 0.962 on the SSDD, HRSID, and SAR-Ship-Dataset, respectively, with a model size of only 14.34M parameters and 44.4 giga floating point operations.

Keywords

• Deformable attention
• detection transformer (DETR)
• information aggregation
• ship detection
• synthetic aperture radar (SAR)