Dynamic fusion and robust downsampling for remote sensing object detection

In remote sensing imagery, complex backgrounds and scale variation make small objects difficult to detect. To address this issue, we propose an SRFD-CAFM (Shallow Robust Feature Downsampling - Convolution and Attention Fusion Module) framework, which is an improved RT-DETR (Real-Time Detection Transformer). It integrates CNN and Transformer to balance local and global information. A dynamic weight allocation (DWA) module adaptively regulates feature contributions, while the Shallow Robust Feature Downsampling (SRFD) module preserves fine-grained shallow information to mitigate small-object feature loss. Experiments on multiple datasets show a 1.6-point mAP50–95 improvement over the baseline, confirming the model’s capability in complex scenes and small object detection.