航空遥感图像(ARSI)飞机检测一直是一个重要且具有挑战性的课题。针对现有ARSI飞机检测方法(ARSIAD)检测目标的边缘模糊、小目标的检测精度低、没有充分利用ARSI的全局上下文信息等问题,提出一种基于多尺度U-Net与Transformer (MSU-Trans)特征融合的ARSIAD方法。通过多尺度卷积模块Inception提取ARSI中多样性目标的分类特征,通过Transformer增强模型的全局语义检测性能,通过特征融合模块整合高层和低层特征,得到航空目标图像完整的边缘和纹理特征。该模型结合多尺度U-Net较强的局部特征提取能力和Transformer较强的全局上下文依存关系提取能力,进而提高MSU-Trans的整体检测性能。在ARSI集上的试验表明,与U-Net、多尺度U-Net、注意力U-Nets相比,MSU-Trans具有较高的检测精度,精度超过95%,该方法为ARSIAD提供一定的技术支撑。

Aerial remote sensing image aircraft detection (ARSIAD) is an important and challenging research. Aiming at the problems of existing ARSIAD methods, such as blurred edges of detection aircrafts, low detection accuracy of small aircrafts and insufficient use of global context information of aerial remote sensing image (ARSI), an ARSIAD method based on feature fusion of multi-scale U-Net and Transformer (MSU-Trans) is proposed. In MSU-Trans, multi-scale convolution module inception is used to extract the classification features of various aircrafts in ARSI, Transformer is used to enhance the global semantic detection performance of the model, and feature fusion module is used to integrate high-level and low-level features to obtain complete edge and texture features of aircraft images, and improve the overall detection performance of MSU-Trans. It integrates the strong local feature extraction capability of multi-scale U-Net and strong global context dependency extraction capability of Transformer to improve the overall detection performance of MSU-Trans. Experiments on an ARSI set show that MSU-Trans has higher detection accuracy than U-Net, multi-scale U-Net and attention U-Nets, with the accuracy over 95%. This method provides some technical support for ARSIAD.