R-Sparse R-CNN: SAR Ship Detection Based on Background-Aware Sparse Learnable Proposals

TL;DR

R-Sparse R-CNN introduces a rotated, sparse-learnable proposal framework for SAR ship detection, extending Sparse R-CNN with orientation and background context. It combines background-aware proposals (BAPs), Dual-Context Pooling (DCP), and an Interaction Module with cross-attention fusion to learn object-background relationships, removing reliance on RPN and NMS. Across SSDD and RSDD-SAR, it achieves state-of-the-art or near-state-of-the-art accuracy, notably improving AP metrics in inshore scenes where clutter is high, while maintaining competitive efficiency. The approach generalizes to large-scale SAR imagery and holds promise for broader oriented-object detection in remote sensing.

Abstract

We introduce R-Sparse R-CNN, a novel pipeline for oriented ship detection in Synthetic Aperture Radar (SAR) images that leverages sparse learnable proposals enriched with background contextual information, termed background-aware proposals (BAPs). The adoption of sparse proposals streamlines the pipeline by eliminating the need for proposal generators and post-processing for overlapping predictions. The proposed BAPs enrich object representation by integrating ship and background features, allowing the model to learn their contextual relationships for more accurate distinction of ships in complex environments. To complement BAPs, we propose Dual-Context Pooling (DCP), a novel strategy that jointly extracts ship and background features in a single unified operation. This unified design improves efficiency by eliminating redundant computation inherent in separate pooling. Moreover, by ensuring that ship and background features are pooled from the same feature map level, DCP provides aligned features that improve contextual relationship learning. Finally, as a core component of contextual relationship learning in R-Sparse R-CNN, we design a dedicated transformer-based Interaction Module. This module interacts pooled ship and background features with corresponding proposal features and models their relationships. Experimental results show that R-Sparse R-CNN delivers outstanding accuracy, surpassing state-of-the-art models by margins of up to 12.8% and 11.9% on SSDD and RSDD-SAR inshore datasets, respectively. These results demonstrate the effectiveness and competitiveness of R-Sparse R-CNN as a robust framework for oriented ship detection in SAR imagery. The code is available at: www.github.com/ka-mirul/R-Sparse-R-CNN.

Figures (13)

• Figure 1: Illustration of ships detected by (a) HBB and (b) OBB detectors in an inshore SAR image scene.
• Figure 2: Comparison of proposals on object detection pipelines: (left) traditional proposals in Faster R-CNN Faster-RCNN-original-paper, (middle) sparse learnable proposals in Sparse R-CNN Sparse-RCNN, and (right) background-aware proposals in R-Sparse R-CNN.
• Figure 3: Architecture comparison between one-stage, two-stage, and anchor-free object detector ((a) and (b) are re-drawn from Faster-RCNN_vs_Anchor-Free).
• Figure 4: Illustration of sparse learnable proposals within Sparse R-CNN. Each proposal consists of proposal box and learnable proposal feature. The proposal feature will interact with pooled RoI feature via its dynamic head to refine the box and predict the corresponding class (re-drawn from Sparse-RCNN).
• Figure 5: Oriented bounding box representation used in proposed R-Sparse R-CNN.