CGCCE-Net:Change-Guided Cross Correlation Enhancement Network for Remote Sensing Building Change Detection

TL;DR

CGCCE-Net is a novel deep learning model that improves remote sensing building change detection by focusing on special colors and semantic relationships between bi-temporal images, achieving superior accuracy.

Contribution

Introduces CGCCE-Net with change-guided residual refinement, global cross correlation, and semantic enhancement modules for improved building change detection.

Findings

Outperforms existing methods on three public datasets.

Effectively detects buildings with special colors.

Enhances semantic understanding of bi-temporal images.

Abstract

Change detection encompasses a variety of task types, and the goal of building change detection (BCD) tasks is to accurately locate buildings and distinguish changed building areas. In recent years, various deep learning-based BCD methods have achieved significant success in detecting difference regions by using different change information enhancement techniques, effectively improving the precision of BCD tasks. To address the issue of BCD with special colors, we propose the change-guided cross correlation enhancement network (CGCCE-Net). We design the change-guided residual refinement (CGRR) Branch, which focuses on extending shallow texture features to multiple scale features obtained from PVT, enabling early attention and acquisition of special colors. Then, channel spatial attention is used in the deep features to achieve independent information enhancement. Additionally, we construct the global cross correlation module (GCCM) to facilitate semantic information interaction between bi-temporal images, establishing building and target recognition relationships between different images. Further semantic feature enhancement is achieved through the semantic cognitive enhancement module (SCEM), and finally, the cross fusion decoder (CFD) is used for change information fusion and image reconstruction. Extensive experiments on three public datasets demonstrate that our CGCCE-Net outperforms mainstream BCD methods with outstanding performance.