ADMNet: adaptive deformable convolution large model combining multi-level progressive fusion for Building Change Detection
Building Change Detection (BCD) based on high-resolution Remote Sensing Images (RSI) simplifies urban surface monitoring. Nevertheless, the mainstream detection methods utilizing traditional convolution and attention mechanisms are often prone to errors due to the loss of edge detail information and...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-01-01
|
| Series: | Geo-spatial Information Science |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/10095020.2024.2448232 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Building Change Detection (BCD) based on high-resolution Remote Sensing Images (RSI) simplifies urban surface monitoring. Nevertheless, the mainstream detection methods utilizing traditional convolution and attention mechanisms are often prone to errors due to the loss of edge detail information and underutilization of global context information. To address these issues, this paper presents a large model, namely ADMNet, which is built on adaptive deformable convolution and is designed to handles various types of building change information. First, we propose a Siamese neural network based on adaptive deformable convolution (ADC) modules. The ADC module incorporates spatial offset parameters into convolutional kernel sampling and mapping weights to capture irregularly varying edge features for local adaptive receptive fields. Second, we utilize a large model semantically driven to enhance model context awareness and construct long-range feature dependencies from multi-scale edge information, which are then integrated with locally adaptive edge structure features to achieve accurate edge localization. Furthermore, we design a Multi-Level Progressive Feature Fusion (MLPFF) module that enhances feature characterization capabilities to ensure internal integrity and improves model detection performance by integrating a priori knowledge from large-model transfer learning. To evaluate the effectiveness and generalizability of ADMNet, we conduct comparative experiments with current mainstream methods on two building datasets, LEVIR-CD and WHU-CD, and a land cover dataset, SYSU-CD. The results show that ADMNet outperforms all comparative methods. The source code is available at https://github.com/spaceYu180/ADMNet. |
|---|---|
| ISSN: | 1009-5020 1993-5153 |