一种注意力卷积与差值特征结合的遥感变化检测模型

A Remote Sensing Change Detection Model Combining Attentional Convolu-tion with Differencing Features

为了有效地提取变化目标,保证变化目标提取的准确性与完整性,文章提出了一种结合了注意力卷积与差异强化结构的三分支网络ACSCN。该网络具有一个变化学习主分支与两个双时相特征学习附属分支,使用包含通道注意力模块和空间注意力模块的注意力卷积模块来代替普通卷积与池化,并利用强化差异模块来强化两个分支的变化差异特征。为了进一步证明ACSCN的性能,我们用不同方法做了对比。实验证明,相比其它方法,ACSCN能更好地检测目标变化,在边界完整性,提取准确性上具有明显的优势,注意力卷积模块与强化差异模块的加入对方法的性能做了有效地提升。

For the effective extraction of change targets and to censure the accuracy and complete-ness of the extracted change targets,a threc-branch nctwork called ACSCN that combines atten-tion convolution with differcnce-cnhancing structures is proposed in this paper.The network consists of a change-learning main branch and two double-temporal-featurc-learning auxiliary branches.It uses an ACB that replaces regular convolution and pooling with attention and spatial attention modules.Additionally,it employs SCB to enhance the change difference features of the two branches.To further demonstrate the performance of ACSCN,we compare it with differ-ent methods.The results indicate that ACSCN performs better in target change detection,partic-ularly in boundary integrity and extraction accuracy,with significant advantages over the other methods.Incorporating ACB and SCB effectively enhances the method's performance.