一种用于遥感图像变化检测的级联跨尺度网络

A Cascade Cross-scale Network for Remote Sensing Image Change Detection

在遥感图像变化检测中,由于深度学习模型没能充分利用多尺度特征,忽略不同尺度间的语义差距,从而导致伪变化;其次,由于成像角度不同、大气环境复杂、季节四季交替等因素也会所引起的伪变化问题,提出一种用于遥感图像变化检测的级联跨尺度网络,该网络设计了一种级联连接将浅层和深层特征图一起馈送到解码器中,以消除不同尺度间的语义差距;提出了一种跨尺度注意力模块来融合浅层和深层特征图中与变化信息一致的语义信息,以提高对伪变化的鲁棒性。在公开的变化检测数据集(LEVIR-CD)上进行了评估。实验结果表明:所提出的方法在性能上明显优于现有的最先进方法。

In remote sensing image change detection,the deep learning model fails to make full use of multi-scale features and ignores the semantic gap between different scales,which leads to pseudo-changes.Different imaging angles,complex atmospheric environments,seasonal alternations and other factors will also cause pseudo changes.To solve the pseudo change problem,a cascade cross-scale network(CCSNet)for remote sensing image change detection is proposed in this paper.We design a cascade structure,which can feed shallow and deep feature maps together into the decoder to eliminate semantic differences between different scales.In addition,a cross-scale attention module(CSAM)is proposed to effectively integrate the consistent semantic information in both shallow and deep feature maps to enhance the robustness against pseudo changes.The proposed method is evaluated on the public change detection dataset(LEVIR-CD),and the experimental results show that the proposed method significantly outperforms the existing state-of-the-art methods in performance.