RARM-YOLO: Remote Sensing Small-Target Detection Model Enhanced by Dual-Branch Region-Aware Refinement Module

Remote sensing image object detection is characterized by dim features of small objects and complex backgrounds. In most networks, simply enhancing small object features may disrupt global consistency and affect detection results. This paper designs a region-aware refinement module to locate the enhanced target semantic representation and suppress background noise interference, and a dual detection branch to fuse underlying details and shallow semantic features at the neck of the feature pyramid, enhancing the model’s ability to focus on very small targets. Experimental results show that the improved model achieves an mAP50% of 76.0% on the VEDAI dataset, which is 10.6% higher than the original YOLOv8s. The mAP50-95% is 46.5%, which is 7% higher. The precision and recall rates are improved by 7.5% and 12.2% respectively. The generalization performance was verified on VisDrone2019 and SODA-A remote sensing datasets, with mAP50% of 47.1% and 73.8% respectively, a model size of only 29.5MB, balancing lightweight and high detection performance. This method provides a technical approach involving the cooperative optimisation of multiple modules for target detection in complex remote sensing scenes, offering significant application value.