基于注意力与多尺度特征遥感图像建筑物分割

Remote sensing image building segmentation based on attention and multi-scale features

为解决图像分割中的边界粘连、小目标分割问题,提出一种遥感图像语义分割模型FAME-Net。在编码器中用ResNet33取代卷积网络,融合空间和通道注意力机制,提高深层次特征和小型建筑物特征提取能力;在中间增设改进的金字塔模块C-ASPP,卷积核锚点引入拉普拉斯算子,增强中心点局部特征,提高建筑物轮廓描述能力;在解码器中融合多尺度特征,设计平均损失函数,有效利用多尺度信息。采用Inria数据集进行性能测试,其结果表明,FAME-Net模型mIoU比U-Net、Link-Net、D-LinkNet、U-Net++模型分别高出8.94%、5.78%、2.47%和2.12%,小目标和边界粘连分割性能优势明显。

To solve the segmentation problems of boundary adhesions and small objects,a semantic segmentation model of remote sensing image buildings FAME-Net was presented.To improve the model’s ability to extract deep-level features and small buil-ding features,the traditional convolutional network was replaced by ResNet33,and also the attention fusion mechanism and multi-scale feature enhancement were merged in the encoder.An improved pyramid module C-ASPP layer was designed between the encoder and the decoder.The Laplacian operator was introduced into the anchor point of the convolution kernel to enhance the local characteristics of the center point and improve the model’s ability to describe the outline of the building.Multi-scale feature fusion was carried out and an average loss function was designed in the decoding stage so as to use multi-scale information effectively.The Inria data set was used to test the remote sensing image segmentation performance of the FAME-Net model.Experimental results show that the mIoU of the FAME-Net model is 8.94%,5.78%,2.47%and 2.12%higher than that of the U-Net,Link-Net,D-LinkNet and U-Net++models respectively.The performance of small target segmentation and boundary adhesion segmentation is significant.