Spatial-frequency domain fusion with global multi-scale pooling for rotated object detection in optical remote sensing

Abstract With the rapid development of computer vision object detection networks, a variety of detection networks based on remote sensing data have been proposed for directional object detection. However, due to the complex backgrounds and significant variations in scale found in remote sensing scenes, directional object detection faces issues of inconsistency between classification and regression, resulting in limitations in existing directional networks regarding object detection and feature extraction. To address these issues, this paper presented SFMP-Net, an optical remote sensing rotated object detection via spatial frequency domain enhancement and multi-scale pooling. First we designed a Spatial-frequency Domain Dynamic Enhancement module. By dynamically weighting the integration of frequency and spatial domains, this module extracts edge texture and semantic information from feature maps, thereby enhancing feature representation capabilities for small targets and complex scenes. Second, to address feature loss during extraction, we proposed a global multi-scale pooling module, implementing attention-guided max pooling to mitigate feature attrition. Finally, we employed a Rotation Adaptive Dynamic Detection Head, utilizing shared convolutions and task-aligned strategies to enhance detection head angle, localization, and classification performance. Using the YOLOv8n model as a baseline, we validated the model’s effectiveness on two public remote sensing datasets, DIOR-R and DOTAV1.0, achieving high accuracy of 80.16% and 87.63% (calculated using mAP50), thereby demonstrating state-of-the-art performance.