DynaNet: A Dynamic Feature Extraction and Multi-Path Attention Fusion Network for Change Detection

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Bi-temporal remote sensing imagery analysis faces challenges in feature fusion and background noise.
• Building change detection requires capturing subtle spatial and semantic dependencies, which existing methods struggle with.

Purpose of the Study:

• To propose DynaNet, a novel network for improved building change detection in remote sensing.
• To address limitations in feature fusion and noise interference in existing change detection techniques.

Main Methods:

• DynaNet utilizes a Dynamic Feature Extractor (DFE) with cross-temporal gating for feature alignment.
• A Contextual Attention Module (CAM) integrates global context to enhance change region discrimination.
• A Multi-Branch Attention Fusion Module (MBAFM) models inter-scale relationships using attention mechanisms.

Main Results:

• DynaNet achieved state-of-the-art performance on the new Inner-CD dataset with an F1-score of 90.92%.
• The method also demonstrated high performance on LEVIR-CD (92.38% F1-score) and WHU-CD (94.35% F1-score).
• Experiments confirmed DynaNet's effectiveness in detecting fine-grained structural changes.

Conclusions:

• DynaNet offers a robust solution for building change detection by effectively handling feature fusion and noise.
• The proposed network architecture and attention mechanisms significantly improve detection accuracy.
• The Inner-CD dataset provides a valuable benchmark for evaluating building change detection algorithms.