A Triple Dynamic Optimization-Based Multi-Scale Object Detection Network

The object detection in high-resolution remote sensing images is affected by the uniqueness of imaging perspectives and the complexity of detection targets. Existing models face core challenges such as feature degradation of small objects, extreme scale variations, and intricate, dynamic backgrounds. To address these issues, this paper proposes a Triple Dynamic Optimization framework for remote sensing target detection (TDO-YOLO). First, construct the Deformable Reparameterized Feature Module (DRFM), which utilizes multi-branch feature fusion and structural reparameterization techniques to achieve refined capture of tiny and multi-scale targets. Second, we design an Adaptive Receptive Field Fusion Module (ARFF) that overcomes fixed receptive field limitations in traditional convolutions through parallel multi-scale kernel extraction and geometry-aware feature reorganization, significantly improving multi-scale object detection accuracy. Third, we propose a Convolutional Dynamic Position-Encoding Residual Attention Mechanism (CPERA), leveraging long-range dependency modeling and background suppression to effectively mitigate complex background interference and reduce false detection rates. Finally, we introduce a Multi-Scale Aware IoU (MSA-IoU) loss function specifically designed for remote sensing images, incorporating multi-scale perception mechanisms and dynamic weighting to enhance detection precision. Experimental results on the RSOD and DIOR datasets demonstrate that TDO-YOLO achieves mAP$_{50}$ scores of 96.6% and 68.7%, representing improvements of 3.87% and 3.15%, respectively, over the baseline model YOLOv11. Compared to current state-of-the-art networks, TDO-YOLO exhibits superior detection accuracy. The complete network model code is available at https://github.com/ChengXCode/TDO-YOLO.