改进的U-Net在建筑物变化检测中的应用

Application of Improved U-Net in Building Change Detection

变化检测是对地观测应用中的一项重要任务。然而现有的基于深度学习的变化检测方法在高分遥感图像建筑物变化检测任务中仍存在变化对象和背景之间分界模糊、小变化目标漏检等问题。针对这些问题,提出了一种基于高斯差分金字塔和注意力特征传递的遥感图像建筑物变化检测方法。该方法采用编码器解码器结构,在编码阶段使用高斯差分金字塔获取双时相遥感图像多尺度边缘特征信息,融合不同尺度下的边缘特征信息,增强图像边缘特征表达能力。在解码部分引入注意力特征传递机制,将高层语义信息与低层建筑物细节有效融合,以捕获特征中的显著信息、抑制无效特征信息,提升小变化目标的检测能力。该方法在公开的LEVIR-CD、WHU-CD数据集上进行训练和测试。实验结果表明,相比于其他同类方法,改进方法在对不同尺度建筑物目标的变化检测中展现了良好的适应性,在保证较低算力消耗的基础上,精确率、召回率、F1和Kappa值方面均有较大提升。

Change detection is an important task in earth observation applications.However,existing deep learning based change detection methods still face problems such as blurred boundaries between changing objects and backgrounds,and missed detection of small changing targets in high-resolution remote sensing image building change detection tasks.A remote sensing image building change detection method based on Gaussian difference pyramid and attention feature transfer is proposed to address these issues.It adopts an encoder decoder structure and uses Gaussian difference pyramid to obtain multi-scale edge feature information of dual temporal remote sensing images during the encoding stage,which integrates edge feature information at different scales to enhance the ability of image edge feature expression.Introducing an attention feature transmission mechanism in the decoding section,effectively integrating high-level semantic information with low-level building details to capture salient information in features,suppress invalid feature information,and improve the detection ability of small change targets.The proposed method is trained and tested on publicly available LEVIR-CD and WHU-CD datasets.The experimental results show that compared with other similar methods,the improved method demonstrates good adaptability in detecting changes in building targets of different scales.While ensuring low computational power consumption,the accuracy,recall,F1,and Kappa values are significantly improved.