A Transformer-based Network with Differential Feature Triple Refinement
for Bitemporal Remote Sensing Image Change Detection
Abstract
In change detection (CD), how to reduce the interferences of pseudo
changes and accurately recognize the change of interest (COI) are two
important challenges. Recently, considering the powerful long-distance
modeling ability of the transformer, some methods try to introduce the
transformer into CD and have already proposed several useful CD
strategies. However, the existing strategies either do not directly work
on the change of interest (COI) or are difficult to give full play to
the advantages of the transformer. Therefore, in this paper, we propose
a new CD strategy to tackle the above challenges. Specifically, we focus
on the difference domain and propose the differential feature triple
refinement strategy to precisely characterize COI. We first adopt a
CNN-based differential feature extraction (DFET) module to extract the
possible detail differences between bitemporal images. Then, we
introduce a transformer-based differential feature enhancement (DFEH)
module to capture and enhance the COI regions from the preliminarily
extracted differences. Finally, we utilize a CNN-based differential
feature fusion (DFFS) module to integrate the fine-grained information
into the enhanced COI regions. Based on the proposed strategy, we design
a new network named DiFormer. We verify six effective hyperparameter
configurations and conduct experiments on four commonly researched CD
datasets. Extensive experiment results indicate that our proposed
strategy has the outstanding generalization ability and obtains the
better balance between computation costs and model performance.
Peculiarly, when only adopting the Natural Scene Image Pretraining
(NSIP), our method still exceeds the recently proposed CD methods which
especially focus on the improvement of Remote Sensing Image Pretraining
(RSIP).