SAS-NET: Similarity Attention Siamese Network for Building Change Detection in UAV Images

Yikui Zhai, Wenba Li, Zijun Tan, Jianhong Zhou, Qing Li, Zilu Ying

Abstract: Change detection refers to extract change information using deep learning or traditional image processing methods to quantitatively analyze and characterize landmark changes on bi-temporal images. Currently, change detection is mainly a pixel-level task, and obtaining accurate change detection segmentation predictions requires a more elaborate and complex model architecture design. To simplify the change detection task, we proposed a novel similarity detection model, Similarity Attention Siamese Network (SAS-NET). It analyzed and predicted if the bi-temporal image patches were similar, and simplified pixel-level change detection tasks to patch-level similarity classification prediction tasks. In this work, a UAV Similarity Detection Dataset (UAV-SD) was also proposed to explore the advantages of patch-level prediction tasks over pixel-level change detection tasks. The proposed method achieved 90.5% accuracy on UAV-SD, which proves that it is more effective than other advanced change detection methods.