基于注意力机制改进的Segformer遥感图像语义分割方法

Improved Segformer network semantic segmentation method for remote sensing images based on attention mechan

针对Segformer处理具有复杂空间和频谱特征的遥感影像时存在局部感受野限制以及深层语义特征损失等问题,提出在Segformer不同层级模块间嵌入不同注意力模块的多级分层编码器网络结构:在Block2之前嵌入极化注意力模块PSA,用以增强网络对大尺度特征的空间感知能力,缓解特征语义损失,并在Block3和Block4之前嵌入高效通道注意力模块ECA获取通道的加权特征,从而增强网络对重要特征的识别能力和感知能力,从终以多特征级联的方式实现像素级遥感影像的语义分割。通过在GID和BCDD数据集上进行测试,与原Segformer相比,新网络在两个数据集的mIOU(%)分别提高了1.85%和1.63%。

In view of the problems of local Receptive field limitation and deep semantic feature loss when Segformer processes remote sensing images with complex spatial and spectral characteristics,a multi-level layered encoder network structure with different attention modules embedded between Segformer modules at different levels is proposed:polarization attention module PSA is embedded before Block2,To enhance the network's spatial perception of largescale features,alleviate feature semantic loss,and embed efficient channel attention module ECA before Block3 and Block4 to obtain weighted features of the channel,thereby enhancing the network's recognition and perception ability of important features,and ultimately achieving pixel level semantic segmentation of remote sensing images through multiple feature cascades.Through testing on the GID and BCDD datasets,compared to the original Segformer,the new network has increased mIOU(%)by 1.85%and 1.63%,respectively,in both datasets.