We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
TChange: A Hybrid Transformer-CNN Change Detection Network.
- Authors
Deng, Yupeng; Meng, Yu; Chen, Jingbo; Yue, Anzhi; Liu, Diyou; Chen, Jing
- Abstract
Change detection is employed to identify regions of change between two different time phases. Presently, the CNN-based change detection algorithm is the mainstream direction of change detection. However, there are two challenges in current change detection methods: (1) the intrascale problem: CNN-based change detection algorithms, due to the local receptive field limitation, can only fuse pairwise characteristics in a local range within a single scale, causing incomplete detection of large-scale targets. (2) The interscale problem: Current algorithms generally fuse layer by layer for interscale communication, with one-way flow of information and long propagation links, which are prone to information loss, making it difficult to take into account both large targets and small targets. To address the above issues, a hybrid transformer–CNN change detection network (TChange) for very-high-spatial-resolution (VHR) remote sensing images is proposed. (1) Change multihead self-attention (Change MSA) is built for global intrascale information exchange of spatial features and channel characteristics. (2) An interscale transformer module (ISTM) is proposed to perform direct interscale information exchange. To address the problem that the transformer tends to lose high-frequency features, the use of deep edge supervision is proposed to replace the commonly utilized depth supervision. TChange achieves state-of-the-art scores on the WUH-CD and LEVIR-CD open-source datasets. Furthermore, to validate the effectiveness of Change MSA and the ISTM proposed by TChange, we construct a change detection dataset, TZ-CD, that covers an area of 900 km 2 and contains numerous large targets and weak change targets.
- Subjects
CONVOLUTIONAL neural networks; INFORMATION sharing
- Publication
Remote Sensing, 2023, Vol 15, Issue 5, p1219
- ISSN
2072-4292
- Publication type
Article
- DOI
10.3390/rs15051219