DFPF-Net: Dynamically Focused Progressive Fusion Network for Remote Sensing Change Detection

TL;DR

The dynamically focused progressive fusion network (DFPF-Net) is proposed, which employs attention mechanisms and edge detection algorithms to mitigate noise interference across varying ranges and outperforms mainstream CD methods.

Abstract

Change detection (CD) has extensive applications and is a crucial method for identifying and localizing target changes. In recent years, various CD methods represented by convolutional neural network (CNN) and transformer have achieved significant success in effectively detecting difference areas in bi-temporal remote sensing images. However, CNN still exhibit limitations in local feature extraction when confronted with pseudo changes caused by different object types across global scales. Although transformers can effectively detect true change regions due to their long-range dependencies, the shadows cast by buildings under varying lighting conditions can introduce localized noise in these areas. To address these challenges, we propose the dynamically focused progressive fusion network (DFPF-Net) to simultaneously tackle global and local noise influences. On one hand, we utilize a pyramid vision transformer (PVT) as a weight-shared siamese network to implement change detection, efficiently fusing multi-level features extracted from the pyramid structure through a residual based progressive enhanced fusion module (PEFM). On the other hand, we propose the dynamic change focus module (DCFM), which employs attention mechanisms and edge detection algorithms to mitigate noise interference across varying ranges. Extensive experiments on four datasets demonstrate that DFPF-Net outperforms mainstream CD methods.

Figures (12)

• Figure 1: Illustrates an overview of the noise effects in CD. Subfigures (a), (b), and (c) show the bi-temporal RS images and the ground truth images after CD. Specifically, the comparison between (a) and (b) highlights the pseudo-change noise from different types of objects, such as trees and buildings in unchanged areas, while the comparison between (b) and (c) emphasizes the shadow noise from buildings in the areas of truly change.
• Figure 2: Provides an overview of the DFPF-Net. To process bi-temporal images, we first employ PVT to construct a weight-sharing siamese structure as the encoder. Then, we input the multi-scale feature information into the PEFM. Subsequently, we propose the DCFM to enhance the weights of the change regions, and finally, we perform differential information fusion using a cross scale interaction decoder.
• Figure 3: Shows the architecture of PEFM. It jointly processes the differential images and bi-temporal images, with shallow and deep features extracted based on residual structure. This progressive approach enhances the image fusion effect, thereby reducing the impact of pseudo-changes and shadow noise.
• Figure 4: The overall architecture diagram of the PVT encoder.
• Figure 5: Shows the architecture of DCFM. In images (a) and (b), the red, purple, and green stars represent pseudo-change areas in the bi-temporal images, while the blue stars indicate real change areas. The agent attention mechanism focuses on the change regions and distinguishes pseudo-changes, while also addressing the shadows produced by buildings. The incorporation of edge detection algorithms further suppresses shadow noise from the buildings.