Remote sensing image rotation target detection algorithm based on improved YOLO26n-obb

Abstract

In the task of rotating target detection in remote sensing images, there are problems such as large scale change, arbitrary direction distribution, and vulnerability to complex background, occlusion and dense arrangement. The existing rotating target detection models usually have problems such as large number of parameters, high computational redundancy, and high deployment costs, which are difficult to meet the lightweight and real-time application requirements in remote sensing scenarios. Therefore, this paper proposes a lightweight remote sensing image rotation target detection algorithm based on improved YOLO26n-obb. A Dynamic Multi-scale Direction- Aware Module ( DMSDA ) is designed to improve the C3K2 module, which enhances the feature expression ability of different scale targets and arbitrary direction targets while maintaining the lightweight of the network. Then, a C2 Efficient Multi-scale Dynamic module ( C2EMD ) is designed to replace the C2 PSA module in the original network, which further improves the robustness of the network to multi-scale, large aspect ratio and rotating targets. The experimental results show that the mAP50 and mAP50-95 values of the improved algorithm on the DOTA subset are increased by 3.4 and 3.3 percentage points respectively, and the lightweight degree of the algorithm is improved slightly, which proves the effectiveness of the improved algorithm.