面向建筑物变化检测的主体边缘分解与重组神经网络

Main body, edge decomposition and reorganization network for building change detection

针对建筑物变化检测中传统神经网络方法易产生锯齿形边界,难以准确识别密集建筑物区域变化地物边界的问题,本文提出一种基于主体、边缘分解与重组网络的变化检测方法。该方法基于地物主体内部像元间相似性较强,而边缘像元间相似性较弱的特点,构建对地物主体与边缘分别建模的变化检测模型。该模型首先通过孪生残差网络结构提取双时相多尺度差值特征:然后,利用可学习的流域分离出地物的主体特征和边缘特征;其次,设计特征优化结构,利用主体标签和边缘标签对分离后的主体特征和边缘特征进行精准优化;最后,将优化后的主体特征和边缘特征进行重组,形成端到端的变化检测模型。通过在公开的建筑物数据集LEVIR-CD进行试验,结果表明,相较于基于U-Net网络的方法和结合时空注意力的方法,该方法能够准确识别变化建筑物的边界,获得更优的检测结果。

Traditional neural network methods for building change detection tend to produce the saw-tooth boundaries, and they are difficult to accurately identify change boundaries in dense building areas. To address that, this paper proposes a change detection method based on main body, edge decomposition and reorganization network. The proposed method performs change detection by respectively modeling body and edges features of buildings, which is on the basis on the characteristics of strong similarity between the body pixels and weak similarity between the edge pixels. In the definition of the proposed method, we first yield dual-temporal multi-scale difference features using a Siamese ResNet structure, and then separate the body features and edge features of buildings by learning a flow field. Subsequently, a feature optimization structure is designed to refine the body and edge features using the body and edge tags. Finally, the optimized body and edge features are reorganized to generate an end-to-end change detection model. Experiments have been performed by using the publicly available building dataset LEVIR-CD, and the results show that the proposed method can accurately identify the boundaries of changing buildings, and obtain better results compared with the methods based on U-Net network and these combining spatial-temporal attention.