Bi-temporal Gaussian Feature Dependency Guided Change Detection in Remote Sensing Images

TL;DR

The BGFD has effectively reduced pseudo changes and enhanced the detection capability of detail information, and a novel detail feature compensation (DFC) module is designed, which compensates for detail feature loss and contamination introduced during the upsampling process from the perspectives of enhancing local features and refining global features.

Abstract

Change Detection (CD) enables the identification of alterations between images of the same area captured at different times. However, existing CD methods still struggle to address pseudo changes resulting from domain information differences in multi-temporal images and instances of detail errors caused by the loss and contamination of detail features during the upsampling process in the network. To address this, we propose a bi-temporal Gaussian distribution feature-dependent network (BGFD). Specifically, we first introduce the Gaussian noise domain disturbance (GNDD) module, which approximates distribution using image statistical features to characterize domain information, samples noise to perturb the network for learning redundant domain information, addressing domain information differences from a more fundamental perspective. Additionally, within the feature dependency facilitation (FDF) module, we integrate a novel mutual information difference loss ($L_{MI}$) and more sophisticated attention mechanisms to enhance the capabilities of the network, ensuring the acquisition of essential domain information. Subsequently, we have designed a novel detail feature compensation (DFC) module, which compensates for detail feature loss and contamination introduced during the upsampling process from the perspectives of enhancing local features and refining global features. The BGFD has effectively reduced pseudo changes and enhanced the detection capability of detail information. It has also achieved state-of-the-art performance on four publicly available datasets - DSIFN-CD, SYSU-CD, LEVIR-CD, and S2Looking, surpassing baseline models by +8.58%, +1.28%, +0.31%, and +3.76% respectively, in terms of the F1-Score metric.

Figures (17)

• Figure 1: The images show the bi-temporal images and the detection results of ChangeCLIP respectively, with green indicating false positive (FP) and red indicating false negative (FN). The top set of images illustrates pseudo changes resulting from illumination difference, while the bottom set of images demonstrates the difficulty in preserving detail of buildings.
• Figure 2: The line graph illustrating the comparative experimental results on the DSIFN-CD dataset. Our model has been distinctly highlighted in red and enlarged for emphasis. It can be observed that our model achieves the highest position across all five metrics.
• Figure 3: The overall architecture of BGFD. Our innovative modules consist of GNDD, DFC, and FDF module composed jointly by GAM and $L_{MI}$. The network is composed of four main parts. The feature extraction part includes image encoders and text encoders for extracting visual and semantic features of the image, respectively. The Gaussian noise domain disturbance part processes the visual features by approximating distributions and sampling from it to eliminate redundant domain information. The difference image extractor generates three types of difference images, with the concatenated one being processed by the FPN structure, utilizing the DFC module to compensate the detail information. Lastly, A decoder is used to handle semantic features and multiple difference features.
• Figure 4: Structure of DFC. The DFC module consists of two branches. The GEM branch is used to further refine the global information in small-scale feature maps, while the DEM is responsible for enhancing the detail information in small-scale feature maps.
• Figure 5: Structure of GNDD. The $\mu$, $\sigma$ refer to the mean value and standard deviation of feature map. In the figure, we only illustrate the computation process for feature maps of time A. In practice, we similarly perform calculations on the multi-scale feature maps of time B.