摘要： 遥感图像变化检测是识别同一区域前后时段两张图像之间像素级变化，能够精确判断目标区域状态变化。现有的遥感图像变化检测方法是在不同时间流中引入注意力机制来强化变化区域图像特征，并将其叠加以实现特征融合，不能有效地挖掘与应用不同时间流特征之间的关系。基于特征提取网络，提出一种在时间维度上基于像素位置偏移的图像特征差异增强方法。该方法可学习不同时相图像特征之间对应区域的像素变化偏移量，增强单时相特征图中发生变化区域和无关区域之间的特征差异。在此基础上，构建一个针对光学遥感图像中建筑物变化的检测框架，以ResNet18网络和多层感知机结构分别作为编码器、解码器，在LEVIR-CD、LEVIR-CD+和S2Looking 3个公开数据集上进行实验，结果表明，基于特征增强的图像变化检测方法的F1值分别为90.74%、86.11%和62.25%，相比目前最优的BIT方法分别提高了1.43、3.31和0.4个百分点。

关键词: 遥感图像, 变化检测, 差异增强, 深度学习, 语义特征

Abstract: Remote sensing image change detection is used to identify the pixel-level changes between two images before and after the same area and accurately judge the state change of the target area.The conventional method, which introduces attention mechanism into different time streams and then superimposes them together to achieve feature fusion, cannot effectively extract and apply the relationship between different time flow features.This paper presents the introduction of an image feature difference enhancement method, based on pixel position offset in time dimension, into the feature extraction network.Thus, the proposed method learns the offset of pixel change in the corresponding region among different phase features.Subsequently, it enhances the feature differences between changed and unrelated regions in the single-phase feature map.Furthermore, this paper proposes a building change detection framework that employs ResNet18 network and Multilayer Perceptron(MLP) structure as the encoder and decoder, respectively.The experimental results on three public data sets-LEVIR-CD, LEVIR-CD+and S2Looking-show that the proposed method achieves 90.74%, 86.11% and 62.25%, respectively, in terms of F1 value.Compared with those achieved for the current optimal BIT method, these values are higher by 1.43, 3.31, and 0.4 percentage points, respectively.

Key words: remote sensing image, change detection, differences enhanced, deep learning, semantic feature

中图分类号: 

• TP391