基于渐进生长Transformer Unet的遥感图像建筑物分割

Building Segmentation in Remote Sensing Image Based on Progressive Growing Transformer Unet

针对深度卷积神经网络在遥感图像地物分割任务中存在树木遮挡目标无法提取、小目标细节丢失等问题,提出了一种基于渐进生长机制的Transformer Unet(PGT-Unet)卷积神经网络模型,通过渐进式逐步增加输入图像的分辨率和模型的深度,使得模型首先在尺度较小的图像获得比较收敛的输出结果,之后在每个阶段可以集中注意力学习相邻尺度的特征信息,最终在原尺度图像上得到最后的建筑物分割结果;同时,在网络模型的编码阶段和解码阶段中引入了Transformer Block模块进行特征提取和特征融合,获得更大的感受野和更强的上下文塑造能力,达到提升模型灵敏度和精确度的目的。在Inria Aerial Image Labeling数据集的建筑物遥感图像上进行实验,结果表明,所提模型能够自适应学习到不同大小目标、遮挡目标的丰富细节特征,从而提升建筑物分割精度,分割结果的平均交并比(Intersection over Union, IoU)为0.775。

In remote sensing image segmentation tasks, deep convolutional neural networks are not able to extract targets blocked by trees and are facing the problem of missing small target details. Thus, a Progressively Growing Transformer Unet(PGT-Unet) convolutional neural network model is proposed. By gradually increasing the resolution of the input image and the depth of model, the model firstly obtains a relatively convergent output result on a smaller-scale image and then focuses on learning feature information of adjacent scales at each stage. Finally, the eventual building segmentation result is obtained on the basis of original scale image. In addition, the Transformer Block module is introduced into coding and decoding stages of the network model for feature extraction and feature fusion so as to obtain a broader receptive field and stronger context shaping ability, thus achieving the purpose of improving sensitivity and accuracy of the model. This method is tested on the remote sensing images of buildings in Inria Aerial Image Labeling dataset. Experimental results show that the proposed model can adaptively learn rich details of different size of targets and blocked targets, thereby improving the accuracy of building segmentation. The experimental segmentation results’ average Intersection over Union(IoU) ratio is 0.775.