RAF-Net: Residual Attention Fusion for Object Detection in Remote Sensing Images

Since remote sensing images are characterized by complex background, small and dense target pixel share, and large scale variation differences, this paper proposes a target detection algorithm for remote sensing images with residual attention fusion. Firstly, the GAM (Global Attention Mechanism) module is utilized to aggregate global context information at the tail of the backbone network; secondly, the RAF (Residual Attention Fusion) module is designed to aggregate local context information at the tail of the feature aggregation network to retain the salient part of the feature information and to improve the detection effect of the target. Experimental results on the RSOD dataset show that the proposed algorithm achieves an average accuracy mean of 97.5%, which is 13.9% better than the benchmark model YOLOv5m, while the model meets the real-time requirements.