基于改进U-Net模型的航空影像建筑物变化检测

Aerial Image Building Change Detection Method Based on Improved U-Net Model

针对航空影像建筑物变化检测中存在的小目标变化类易被漏检和摄影倾角导致场景分割较为粗糙等问题,本文提出了一种基于改进U-Net模型的航空影像建筑物变化检测方法。首先,将注意力机制引入U-Net网络,从而放大变化类的像素特征来提高提取结果精度;其次,使用CRFs全连接条件随机场,对初步变化检测结果进行后处理,消除摄影倾角产生的阴影问题,优化建筑物边界轮廓。在WHU建筑物变化检测数据集上的实验结果表明,在引入注意力机制和CRFs全连接条件随机场后,建筑物变化检测结果的准确率、召回率、F1值和总体精度4项指标有了明显提升,分别达到0.884、0.870、0.950、0.859,均优于传统U-Net模型。

In order to solve the problems of building change detection in aerial images,such as small target change is easy to be missed and the scene segmentation is rough due to the angle of photography,this paper proposes an improved U-Net model for building change detection in aerial images.Firstly,the attention mechanism is introduced into The U-Net network to amplify the pixel features of the change class to improve the accuracy of the extraction results.Secondly,the CRFs fully connected conditional random field was used to post-process the preliminary change detection results to eliminate the shadow problem caused by the photographic angle and optimize the building boundary contour.Experimental results on WHU building change detection dataset show that the accuracy,recall rate,F1 value and overall accuracy of building change detection results are significantly improved after the introduction of attention mechanism and CRFs fully connected condition random field,reaching 0.884,0.870,0.950 and 0.859,respectively.They are su-perior to the traditional U-Net model.