The paper proposes a Siamese-based framework, called ChangeBind, for remote sensing change detection. The key contributions are:
The framework introduces a hybrid change encoder that combines the benefits of convolutional operations and self-attention mechanisms to capture both subtle and large change regions effectively.
The change encoder utilizes multi-scale features extracted from a Siamese-based ResNet backbone to encode change information at different levels of granularity.
The convolutional change encodings (CCE) capture fine-grained textural details, while the attentional change encodings (ACE) focus on learning global contextual representations. These complementary encodings are fused to obtain rich change representations.
Extensive experiments on two challenging change detection datasets, LEVIR-CD and CDD-CD, demonstrate the superiority of the proposed approach over state-of-the-art methods, achieving new benchmarks in terms of F1-score, IoU, and overall accuracy.
The qualitative results show that the hybrid change encoder can better detect both subtle and large-scale changes compared to existing CNN-based and transformer-based approaches.
To Another Language
from source content
arxiv.org
Viktige innsikter hentet fra
by Mubashir Nom... klokken arxiv.org 04-29-2024
https://arxiv.org/pdf/2404.17565.pdfDypere Spørsmål