基于结构一致性能量模型的异源光学和SAR图像变化检测

A structure consistency-based energy model for heterogeneous optical and SAR images change detection

异源遥感图像变化检测是遥感领域越来越受到关注的课题.由于成像机制的不同,无法通过直接比较不同源的图像来检测变化.为了克服这一难题,本文利用与成像模态无关的结构一致性建立异源图像之间的联系.首先,通过构造K近邻图来表征图像结构,并挖掘图像结构关系与节点变化状态之间的关联.而后,提出了基于图像结构一致性的能量模型来刻画这种关联,并设计了两个能量函数分别度量节点发生变化和未发生变化时的能量损失.最终,通过求解该能量模型得到每个节点发生变化的概率,并对其进行分割得到变化检测结果.由于本文所提出的方法是基于图像结构的,所以它对图像噪声、成像条件等因素不敏感,同时由于该方法对节点状态直接建模,能充分利用能量模型的灵活性,所以它可适用于不同的任务场景.在5个实际数据集上的实验结果验证了本文方法能够取得更高效、更准确的检测结果.

Heterogenous change detection is a topic of increasing interest in remote sensing.Due to the different imaging mechanisms,it is impossible to compare the multitemporal images from different sources directly.To address this challenge,this paper uses structure consistency to establish the connection between heterogeneous images,which is imaging modality-invariant.First,we construct a K-nearest neighbor graph to characterize the image structure and explore the relationship between the image structure and the node state.Then,we propose a structure consistency-based energy model to characterize these relationships and construct two functions to measure the energy loss both when the change occurs and when it does not occur.Finally,we solve the energy model to obtain the change probability of each node and segment them to obtain the change detection results.Since the proposed method is based on the image structure,it is insensitive to image noise and imaging conditions.Meanwhile,the method models the node states directly and can fully exploit the flexibility of the energy model,enabling it to be applied to different change detection scenarios.The experimental results on five real datasets verify that the proposed method can achieve more efficient and accurate results.