遥感图像变化检测是遥感领域的重要研究方向.针对多尺度编码特征语义差异性和空间信息差异性引入伪变化干扰的问题,本文提出了一种两阶段特征金字塔的变化检测网络TS-FPCDN(Two Stage-Feature Pyramid based Change Detection Network),通过两阶段解码器增强变化特征描述,抑制伪变化信息干扰.首先,采用孪生编码网络对双时相遥感影像进行编码特征提取,并生成多尺度初始差异特征.由于初始差异特征中存在大量噪声和伪变化信息,通过第一阶段特征金字塔结构和双注意力引导的多尺度差异特征融合机制,进行多尺度差异特征语义信息和空间信息的交互,缓解多尺度特征语义的差异性和空间信息的差异性,初步去除伪变化信息干扰,生成多尺度初始变化特征.然后,为进一步提高变化特征描述和去除伪变化,设计了第二阶段特征金字塔,对多尺度变化特征逐层重优化,并进行变化预测.最后,在两个变化检测数据集LEVIR-CD和WHU-CD上开展了一系列实验,实验结果证明了本文提出方法的有效性.
Two Stage-Feature Pyramid Based Remote Sensing Images Change Detection
Change detection in remote sensing is a significant research focus within the field.This paper proposes a two-stage feature pyramid-based change detection network to address the challenges of pseudo-change and noise caused by semantic and spatial differences in multi-level features extracted by the encoder.The two-stage decoder was used to en-hance the representation of the change feature and suppress the information interference of pseudo-change.First,the Sia-mese encoder network was used for bi-temporal remote sensing image encoding,feature extraction,and multi-scale ini-tial difference feature extraction.Given the presence of excessive noise and pseudo-change information in the initial dif-ference feature,a first-stage feature pyramid structure and a spatial-channel dual attention fusion mechanism were pro-posed to facilitate the interaction of semantic information and spatial information in the multi-scale difference feature,re-lieve the semantic difference and spatial difference of the multi-level feature,initially remove the pseudo-change infor-mation interference,and generate a multi-scale initial change feature.Subsequently,to further improve the representa-tion of the change feature and remove the pseudo-change,the second-stage feature pyramid structure was proposed to optimize the multi-scale change feature stage by stage and then predict change detection.Finally,a series of experiments were conducted on two change detection datasets,LEVIR-CD and WHU-CD,and the experimental results proved the effectiveness of the proposed method.
NETL
NSTL
万方数据
