RS-YOLO: A Lightweight Improved Framework of YOLOv12 for Multi-Target Recognition in Remote Sensing Images

In high-resolution remote sensing images, the significant scale differences among targets, complex backgrounds, and dense distribution of small targets pose severe challenges to traditional detection models. To address this, this paper proposes RS-YOLO, a lightweight multi-target detection model specifically designed for remote sensing scenarios. Firstly, by optimizing the backbone network architecture, we introduced and enhanced the Dynamic Tanh (DyT) normalization mechanism. Given the characteristic of extreme value density in remote sensing images, DyT achieves activation scaling through dynamic learning rate adjustment and channel-independent parameter design, thereby avoiding the statistical dependency issues associated with traditional normalization. Secondly, we optimized and introduced an enhanced IoU loss function (Inner-IoU), which integrates center prior, rotation angle weighting factors, and small target weighting mechanisms to weight-correct IoU. Experimental results on the DOTA-v1.0 dataset show that RS-YOLO achieves $78.9 \% \mathrm{mAP}_{50}$ on DOTA with only 2.52 M parameters and 5.9 G floating-point operations (FLOPs), and an inference latency of $\mathbf{3 5. 8}$ milliseconds, fully demonstrating its lightweight advantage.Notably, the model’s mAP<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</inf> exceeds $80 \%$ in the detection tasks of multiple typical targets. Significant performance improvements are particularly evident in difficult cases such as extremely small targets and highrotation targets. RS-YOLO provides an efficient and feasible solution for rapid multi-target inspection in remote sensing images.