A Framework for Lightweight Object Detection in Remote Sensing Images

This Remote sensing image detection is a key technology in Earth observation and has important uses in astronomy, military, and agriculture. However, current remote sensing images often have small targets, complex backgrounds, and serious occlusion. To address these issues, we propose a new detection algorithm called YOLOv8n-tubo. This method makes the model more accurate and lighter, which helps handle the complexity and variety of remote sensing images. First, we present BMFPN, a new backbone network structure for the feature pyramid. It boosts the model’s ability to detect large-scale objects. Next, we improve the linear attention mechanism using the Mamba concept. We replace the forget gate with positional encoding (RoPE) so that the model keeps parallel computing and fast inference. Finally, we use ODConv (Omni-Dimensional Dynamic Convolution) instead of c2f. This reduces the model’s computational complexity while improving its accuracy. Our experiments on the DOTAv1.0 dataset show that YOLOv8n-tubo raises precision by 2.5%, recall by 0.9%, and mAP by 1.2% over the original YOLOv8. Meanwhile, GFLOPs drop by 0.8. Through these optimizations, our model offers both speed and accuracy for detecting small targets in remote sensing images, providing strong support for related applications.