摘要
目前深度学习方法在遥感影像变化检测方面取得了较大的进步,然而现有的遥感影像变化检测方法仍然以全监督网络为主,其网络性能严重依赖标签数据的数量和质量。为此,提出了一种基于均值教师模型联合多级扰动的半监督遥感影像变化检测网络(UniMTCD-Net)。首先,将不同性质的强扰动分离到不同的分支分别进行学习并约束一致性,形成多样化的扰动空间,避免了单分支学习困难的问题,从而有效提升对无标签数据的利用效率;其次,采用均值教师模型,不仅扩展了教师模型生成的伪标签和学生模型输出的强预测之间的差异,同时教师模型参数通过指数移动平均(EMA)更新的方式,使得伪标签的生成更加准确。实验结果表明,与主流半监督方法相比,UniMTCD-Net具有更好的检测性能,尤其在5%的标签训练数据下检测性能更加优秀,进一步验证了UniMTCD-Net在遥感影像变化检测中的有效性和优越性。
Currently,deep learning methods have made significant progress in remote sensing image change detection.However,the existing approaches in remote sensing image change detection primarily rely on fully supervised networks,which heavily depend on the quantity and quality of label data.In order to solve these problems,this paper proposes UniMTCD-Net,a semi-supervised remote sensing image change detection network that combines the mean teacher model with multi-level perturbation.Firstly,different types of strong perturbations are separated into different branches for learning and consistency is constrained to form a diversified perturbation space,avoiding the difficulty of single branch learning and effectively improving the utilization efficiency of unlabeled data.Secondly,by using the mean teacher model,not only the difference between the pseudo labels generated by the teacher model and the strong predictions output by the student model are extended,but also the teacher model parameters are updated by exponential moving average(EMA),making the generation of pseudo labels more accurate.Experimental results demonstrate that compared with mainstream semi-supervised methods,UniMTCD-Net has better detection performance,especially under 5%labeled training data,the detection performance is more superior,further verifying the effectiveness and superiority of UniMTCD-Net in remote sensing image change detection.
作者
于松岩
翟钰杰
许叶彤
赵伟强
雷涛
YU Songyan;ZHAI Yujie;XU Yetong;ZHAO Weiqiang;LEI Tao(School of Electronic Information and Artificial Intelligence,Shaanxi University of science and technology,Xi'an,Shaanxi 710021,China;Shaanxi Joint Laboratory of Artificial Intelligence,Shaanxi University of Science and Technology,Xi'an,Shaanxi 710021,China;Xi'an Branch,China Electronics Technology Group Corporation Northwest Group Corporation,Xi'an,Shaanxi 710065,China)
出处
《计算技术与自动化》
2024年第4期91-96,共6页
Computing Technology and Automation
基金
国家自然科学基金资助项目(61871259,61861024)
陕西省重点研发计划重点产业创新链项目(2021ZDLGY08-07)。
关键词
变化检测
半监督
一致性
均值教师模型
change detection
semi-supervised
consistency
mean teacher model
