多尺度差分特征增强网络的遥感影像变化检测

Multiscale Difference Feature Enhancement Network for Remote Sensing Image Change Detection

遥感影像变化检测旨在识别不同时期遥感影像中的目标差异,近年来基于卷积神经网络的方法在遥感影像变化检测任务中取得了很大进展。然而,受光照变化和季节更替的影响,不同时期遥感影像中的伪变化问题仍然难以解决。同时,大多方法中未充分利用多尺度特征,导致模型的性能和准确率受到一定程度的限制。针对上述问题,提出一种多尺度差分特征增强的变化检测方法。利用由孪生网络编码器和差分网络编码器组成的并行编码框架分别提取不同层级的特征,将同级的双时特征和差分特征通过拼接的方式建立两者间的互补关系。引入差分特征增强模块获取更具判别性的特征图并将其作为差分网络编码器的补充输入,丰富变化信息的同时增加模型对变化区域的关注度,使其准确地区分地物的真实变化与伪变化。为了增强特征的多样性和表达能力,使用特征错位融合模块实现语义特征的交叉融合,让每个特征中的语义信息得到充分而不同的交互。该方法在CDD数据集和LEVIRCD数据集上的F1分数分别达到了95.45%和92.04%,交并比分别达到了92.26%和82.93%,与其余八种主流方法相比均为最优,实验结果证明了该方法的有效性。

Remote sensing image change detection aims at identifying target differences in remote sensing images of different periods,and methods based on convolutional neural networks have made great progress in remote sensing image change detection tasks in recent years.However,the problem of pseudo-changes in remote sensing images in different periods is still difficult to be solved due to the influence of light changes and seasonal changes.Meanwhile,multi-scale features are not fully utilized in most of the methods,resulting in a certain degree of limitation in the performance and accuracy of the models.To address the above problems,a multi-scale differential feature-enhanced change detection method is proposed.Firstly,a parallel coding framework consisting of a twin network encoder and a differential network encoder is used to extract features at different levels respectively,and the same level of diachronic features and differential features are spliced to establish a complementary relationship between them.Then,the differential feature enhancement module is introduced to obtain more discriminative feature maps as supplementary inputs to the differential network encoder,which enriches the change information and increases the attention of the model to the change area,so that it can accurately distinguish the real change from the pseudo change of the features.Finally,in order to enhance the diversity and expressiveness of the features,the feature mismatch fusion module is used to achieve the cross-fertilisation of the semantic features,so that the semantic information in each feature is fully and differently interacted.The F1 score of this method on the CDD dataset and LEVIR-CD dataset reaches 95.45%and 92.04%respectively,and the intersection over union(IOU)reaches 92.26%and 82.93%respectively,which are optimal compared with the remaining eight mainstream methods,and the experimental results prove the effectiveness of this method.