改进DeepLabV3+的遥感图像建筑物分割

Building segmentation of remote sensing image based on DeepLabV3+ network

针对DeepLabV3+进行遥感图像建筑物分割时存在小目标建筑物漏分、目标建筑物误分以及边界粘合的问题,提出一种改进DeepLabV3+的遥感图像建筑物分割方法。首先,在编码器阶段使用改进的密集空洞金字塔池化DenseASPP模块,获得更大的感受野和更密集的特征金字塔,并行加入条形池化模块,使主干网络有效利用长距离依赖关系。其次,在解码器阶段引入SE通道注意力模块,加强各通道间的关联性,以获取更丰富的边缘特征。最后,将SE模块优化后的特征与原特征进行融合,增强网络的分割性能。在WHU Building数据集上的实验结果表明,本方法的建筑物分割结果在交并比(Iou)和F1指数上分别达到了92.33%和95.54%。

Aiming at the problems of small target building omission,target building misclassification and boundary bonding in remote sensing image building segmentation by DeepLabV3+,this paper proposes an improved remote sensing image building segmentation method for DeepLabV3+.Firstly,an improved dense cavity pyramidal pooling DenseASPP module is used in the encoder stage to obtain larger sensory fields and denser feature pyramids,and a bar pooling module is added in parallel to enable the backbone network to make effective use of the long-range dependencies.Secondly,the SE channel attention module is introduced in the decoder stage to enhance the correlation between channels to obtain richer edge features.Finally,the optimised features from the SE module are fused with the original features to enhance the segmentation performance of the network.The experimental results on the WHU Building dataset show that the building segmentation results of this paper's method achieve 92.33% and 95.54% in the intersection and merge ratio(Iou) and F1 index respectively.