基于DETR和改进去噪训练的光学遥感图像多尺度旋转目标检测(特邀)

DETR with Improved DeNoising Training for Multi-Scale Oriented Object Detection in Optical Remote Sensing Images(Invited)

旋转目标检测是遥感图像解译的重要任务之一,存在目标方向任意、小目标密集排列、目标表示引起的角度周期性等典型问题。针对上述问题,提出一种基于DEtection Transformer(DETR)目标检测器和改进去噪训练的旋转目标检测方法,即arbitrary-oriented object detection Transformer with improved deNoising anchor boxes(AO2DINO)。首先,该方法引入一种多尺度旋转可变形注意力模块,将角度信息以旋转矩阵的形式引入注意力权重的计算,提高了模型对旋转目标的适应能力。其次,针对小目标密集排列问题,提出一种自适应的样本分配器,引入旋转交并比和自适应阈值,实现对密集目标更加精确的采样,提升了模型对小目标的检测能力。最后,在模型中引入基于卡尔曼滤波的交并比(KFIoU)作为回归损失,以解决旋转目标表示引起的角度周期性问题。AO2DINO在DOTAv1.0和DIOR-R两个公开数据集上与典型的旋转目标检测方法进行了比较,在DETR系列旋转目标检测方法中检测精度最高,且训练时收敛速度更快,在训练12个epochs时就几乎达到了其他旋转目标检测方法训练36个epochs时的检测效果。

Oriented object detection is one of the important tasks in remote sensing image interpretation,which faces typical problems such as arbitrary object orientation,dense arrangement of small targets,and angular periodicity caused by target representation,thus,this paper proposes a method called arbitrary-oriented object detection Transformer with improved deNoising anchor boxes(AO2DINO) which based on DEtection Transformer(DETR) and improved denoising training.First,a multi-scale rotated deformable attention(MS-RDA) module is proposed.The MS-RDA module introduces the angle information in the form of rotation matrix for the calculation of attention weights,which improves the adaptability of the model to the orientated objects.Second,this paper proposes a self-adaption assigner(SAA),which uses the rotated intersection over union(IoU) and adaptive threshold to accurately separate dense targets,to improve the small targets detection under the dense arrangement scenarios.Finally,the Kalman filtering IoU(KFIoU) is introduced as the regression loss to solve the angular periodicity problem caused by the representation of orientated objects.Our proposed method is compared with the typical oriented bounding box(OBB) methods on two public datasets,DOTAv1.0 and DIOR-R,and the detection accuracy is the highest among the DETR-based OBB methods,and the convergence speed is faster during training,which only needs 12 training epochs to achieve comparable detection accuracy as other methods using 36 training epochs.