PSNet: change detection with prototype similarity

Change detection is a fundamental problem in remote sensing image processing. Due to the great advantages in learning the knowledge representations and the complex relationship from large-scale datasets, deep learning has made great progress in change detection tasks in remote sensing community. However, most of the existing methods based on deep learning for change detection are implemented by learning differences of image pairs directly without paying considerations in influences of unstructured and temporal changes, or called nature changes, such as light and seasonal changes. In this paper, an end-to-end deep learning network for remote sensing image change detection is proposed, aiming to accurately detect the change of regions from a high-resolution image pair by learning prototype similarity, in which the metric learning is used and it is one of the meta-learning methods to learn change prototypes from support image pairs. The similarity between the query image pairs and the change prototypes can be measured by a learnable CNN metric. The experimental results based on the two public change detection datasets of high-resolution satellite images, CDD and BCDD, show that our proposed method performs better than other state-of-the-art change detection methods with an improvement of 3.5% and 0.4%, respectively.