基于多尺度特征融合注意力CNN的遥感图像语义分割方法

Remote sensing image semantic segmentation method based on multi-scale feature fusion attention CNN

针对遥感图像目标尺度变化较大、分割不够准确的问题,提出了一种融合多尺度特征注意力卷积神经网络(CNN)的图像分割方法。该方法基于卷积块注意力模块提出了改进的有效注意力模块(ECBAM)。在通道注意力模块中加入空洞卷积以降低池化操作造成的特征损失,并在通道注意力模块后添加卷积层对通道注意力特征映射进行特征融合。然后,基于ECBAM设计了一种编码解码架构的卷积神经网络模型ECBAMCNN,其中编码器主要由卷积层、ECBAM和空洞空间金字塔池化模块组成,解码器主要由卷积层和ECBAM组成,并且采用跳跃连接将编码阶段的多尺度信息融合到解码器。实验表明,提出的方法无需预训练和后处理,与SegNet等前沿方法相比取得了更好或相近的分割准确度,在DLRSD和WHDLD测试集上的mIoU分别为67.3%和62.0%。

There are large changes in the target scale of remote sensing images and insufficient segmentation accuracy.So an image segmentation method based on multi-scale feature fusion attention convolutional neural network(CNN)is proposed.This method is an improved Effective Convolutional Block Attention Module(ECBAM)based on the convolutional block attention module(ECBAM-CNN).This module adds dilated convolution to the channel attention module to reduce the feature loss caused by pooling operation,with a convolution layer after the channel attention module to perform feature fusion on the channel attention feature map.Then,a convolutional neural network model with an encoding-decoding architecture is designed based on ECBAM.The encoder is mainly composed of convolutional layers,ECBAM and the Atrous Spatial Pyramid Pooling module.The decoder is mainly composed of convolutional layers and ECBAM,and skip connections are used to fuse the multi-scale information of the encoding stage to the decoder.It is not necessary for the proposed method to pretrain and post-process.But better or similar segmentation accuracy can be achieved compared with cutting-edge methods,such as SegNet et al.The mIoU on the DLRSD and WHDLD testing sets are 67.3%and 62.0%respectively.