Abstract: Facing with the challenges of the vehicle detection in remote sensing images, such as complex backgrounds, multi-scale differences, and difficulty in detecting small targets, a detection method GEM_YOLO based on bidirectional multi-scale feature fusion is proposed. There are three main parts in this method: the first one is a globally efficient attention module that is designed as a feature extractor to achieve lightweight and efficient feature extraction, in order to solve the problem of object detection in complex backgrounds. Secondly, a bidirectional multi-scale feature fusion network is proposed as a feature fusion device, which adopts top-down and bottom-up feature fusion strategies to effectively promote information exchange between features at different levels. Finally, the application of an attention based on the dynamic detection head as a predictor enhances the perception of different scales, spatial positions, and tasks, further improving the accuracy and robustness of object detection. Related experiments are conducted on public datasets DIOR and DOTA, whose average accuracy reaches 92.4% and 81.4% that is significantly superior to other mainstream detection methods. Meanwhile, the fewer parameters and computational complexity provide an efficient solution for vehicle detection within the domain of remote sensing image detection.

Key words: remote sensing images, vehicle inspection, multi-scale feature fusion, attention mechanism, dynamic detection head

摘要： 针对遥感图像中车辆检测面临的背景复杂、多尺度差异和小目标难以检测等挑战，提出了一种基于双向多尺度特征融合的检测方法GEM_YOLO。该方法包括三个主要部分：设计了全局高效注意力模块作为特征提取器，实现轻量化和高效率的特征提取，以解决复杂背景下的目标检测问题；提出了双向多尺度特征融合网络作为特征融合器，采用自顶向下和自底向上的特征融合策略，有效促进不同层次特征之间的信息交互；应用基于注意力的动态检测头作为预测器，增强了对不同尺度、空间位置和任务的感知，进一步提升了目标检测的精度和鲁棒性。在公开数据集DIOR和DOTA上进行相关实验，该方法的平均精度均值达到92.4%和81.4%，显著优于其他主流检测方法，同时具有更少的参数量和计算量，为遥感图像检测领域中的车辆检测提供了一种高效解决方案。

关键词: 遥感图像, 车辆检测, 多尺度特征融合, 注意力机制, 动态检测头