Edge-Aware Attention with Shallow-Guided Fusion for Ship Detection in Remote Sensing

Remote sensing ship detection is of great significance in marine traffic monitoring, port management, national defense security and other fields. Due to the small target scale, complex backgrounds and sparse distribution of ships in remote sensing images, existing real-time detectors still face challenges in localization accuracy and feature utilization efficiency. Taking RT-DETR-s as the baseline, this paper proposes two encoder-level improvements to boost the performance of ship detection in remote sensing imagery. On the attention mechanism side, an Edge-Aware Self-Attention (EASA) module is proposed as the principal methodological contribution. The Scharr edge map is injected into the attention matrix before softmax with a learnable intensity, providing a structural prior that enhances attention interactions between positions with strong edge responses, including ship contours, coastlines, and wake boundaries. The semantic distinction between ship-relevant and irrelevant edges is learned through the query-key projections in the attention mechanism itself, rather than through the edge bias. In terms of feature fusion, a Shallow-Guided Deep Fusion (SGDF) module is designed as an efficient integration strategy of existing techniques (CBAM attention and cross-scale fusion), with its novelty in the attention-first, downsampling-later pipeline that preserves fine-grained spatial details for small-object detection. It injects high-resolution small-object context from the backbone’s shallow layers into deep features, effectively mitigating the loss of spatial information of small ships in deep feature maps via spatial-channel attention. Finally, the existing WIoUv3 loss function is adopted and systematically compared with six other IoU variants to identify the most suitable regression loss for remote sensing ship detection. Experiments demonstrate that compared with RT-DETR-s, the proposed method achieves 2.8% higher mAP50–95 and 3.2% higher APsmall (APs) while maintaining real-time inference speed (37.6 FPS).