Enhanced Change Detection in Unregistered Images With CNNs and Attention GANs

Jinpeng Du, Xiren Zhou, Huanhuan ChenIEEE GRSL 2024 · OpenReview

Abstract

Change detection (CD) involves comparing multitemporal images captured at different times. However, the existing CD methods mainly focus on registered images, neglecting the challenges posed by unregistered image pairs. Besides, the lack of explicit training in distinguishing changes from unchanged areas results in noisy outcomes. To this end, we propose to leverage the convolutional neural network (CNN) and generative adversarial network (GAN) to automatically extract the common region from unregistered images and conduct CD in the extracted region. Specifically, our method utilizes CNN to extract high-level convolutional information from unregistered images. Subsequently, feature matching is employed to identify common regions between the two images. These extracted regions serve as input for the supporting expansion strategy, enabling the creation of a training set for the subsequent training of GAN, which consists of a generator incorporating the channel attention module (CAM) and a discriminator. The optimized generator then produces multiple enhanced coregistered images, finally compared to generate the change map. The experimental results show that for the changed scene image pairs, the proposed method is 4.46% higher than the best performance of the comparison methods on overall accuracy (OA), 2.65% higher on precision (Pre), and 2.3% higher on $F1$ score (F1). For the unchanged scene image pairs, the OA is improved by 2.72%. These demonstrate the superior performance of our proposed method in CD, especially on unregistered image pairs.
BibTeX
@article{du2024enhanced,
  author = {Jinpeng Du and Xiren Zhou and Huanhuan Chen},
  title = {{Enhanced Change Detection in Unregistered Images With CNNs and Attention GANs}},
  journal = {IEEE Geosci. Remote. Sens. Lett.},
  year = {2024},
  url = {https://openreview.net/forum?id=FNw82MYAbS},
}

Loading related content...

Figure