针对遥感图像中的有向物体检测任务,提出了一种基于半监督学习的密集区域卷积神经网络(Dense Region Convolutional Neural Network,D-RCNN)框架,以减少对大规模标注数据的依赖并提高检测精度.在该框架中,利用教师-学生模型通过稠密伪标签生成与一致性损失进行训练,结合伪标签学习与数据扰动,提升模型对无标注数据的有效利用率.针对长尾分布问题,引入了 Seesaw Loss以动态调整各类别权重,进一步优化模型性能.在DOTA数据集上进行的实验表明,D-RCNN在1％、2％、5％标注率下的检测精度AP5.分别较完全监督方法提升了 7.21％、8.02％和2.84％.在低标注率条件下,D-RCNN在多个主要类别上表现出显著的性能优势,验证了其在遥感场景下的有效性.
Semi-supervised Two-stage Remote Sensing Object Detection Method Based on Dense Teacher
For the task of detecting directed objects in remote sensing images,a semi-supervised-learning-based Dense Region Convolutional Neural Network(D-RCNN)framework is proposed to reduce reliance on large-scale labeled data and improve detection accuracy.In this framework,a teacher-student model is utilized for training through dense pseudo-label generation and consistency loss,and pseudo-label learning is combined with data perturbation to enhance the model's effective utilization of unlabeled data.To address the long-tail distribution problem,Seesaw Loss is introduced to dynamically adjust the class weights,further optimizing the model performance.Experiments conducted on the DOTA dataset show that the D-RCNN improves detection accuracy by 7.21％,8.02％,and 2.84％in terms of AP50 at labeling rates of 1％,2％,and 5％respectively compared to fully-supervised methods.Under low labeling rate conditions,the D-RCNN shows significant performance advantages across multiple major categories,validating its effectiveness in remote sensing scenarios.
semi-supervised learningremote sensing imagesdirected object detectionpseudo-label learningconsistency training
万方数据
维普
