基于稀疏Transformer的遥感旋转目标检测

Sparse Transformer Based Remote Sensing Rotated Object Detection

针对遥感图像目标广邻域稀疏、多邻域聚集、方向多样等特性导致检测难度大的问题,提出了一种基于稀疏Transformer的遥感旋转目标检测方法。首先,所提方法在典型端到端Transformer网络的基础上,根据遥感图像的特性,利用Kmeans算法实现多域聚集,从而更好提取稀疏域下的目标特征;其次,为适配旋转目标的基本属性,在边框生成阶段,利用目标包围框的中心点及边框特征学习的策略高效获取目标回归斜边框;最后,为提升网络对遥感目标的检测率,对网络的损失函数进行了优化。在DOTA和UCASAOD遥感数据集上的实验结果表明,所提方法的平均精度分别为72.87%和90.4%,能很好地适应遥感图像中各类旋转目标的形状与分布特性。

A remote sensing rotating target detection approach based on a sparse Transformer is proposed to address the problem of remote sensing image target detection,which is challenging due to the wide neighborhood sparse,multineighborhood aggregation,and multiple orientations characteristics.First,this method uses the Kmeans clustering algorithm to produce multidomain aggregation,to better extract the target features in the sparse domain,based on the typical endtoend Transformer network,and the characteristics of a remote sensing image.Second,to adapt to the basic characteristics of the rotating target,a learning technique based on the target bounding box’s center point and the frame features is proposed in the frame generation stage,to efficiently obtain the target regression oblique frame.Finally,the network’s loss function is further optimized to improve the detection rate of the remote sensing target.The experimental results on DOTA and UCASAOD remote sensing datasets show that the average accuracy of this technique is 72.87%and 90.4%,respectively;thus indicating that it can adapt effectively to the shape and distribution characteristics of various rotating targets in remote sensing images.