A Dual-Branch Fusion Network Based on Reconstructed Transformer for Building Extraction in Remote Sensing Imagery

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Automatic building extraction is crucial for urban planning, demographics, and disaster assessment.
• Convolutional Neural Networks (CNNs) and transformers are common for semantic segmentation but have limitations in feature extraction.
• CNNs struggle with global features due to fixed structures, while transformers can be computationally redundant and poor at local details.

Purpose of the Study:

• To propose a novel dual-branch fused reconstructive transformer network (DFRTNet) for efficient and accurate building extraction from high-resolution remote sensing images (HRSI).
• To address the limitations of existing CNN and transformer models in capturing both global context and local details.

Main Methods:

• Developed a DFRTNet with a dual-branch encoder for local and global feature extraction.
• Integrated a Local and Global Feature Extraction (LGFE) module, featuring Dynamic Range Attention (DRA) for global features and a Local Feature Extraction (LFE) branch for fine-grained details.
• Utilized a Multilayer Perceptron (MLP) for feature fusion and a Channel Attention Module (CAM) in the decoder for enhanced feature representation.

Main Results:

• DFRTNet demonstrated superior performance in building segmentation compared to existing state-of-the-art methods.
• Achieved the best segmentation accuracy on the WHU and Massachusetts building datasets.
• The proposed network effectively balances global context modeling and local feature extraction.

Conclusions:

• DFRTNet offers an efficient and accurate solution for automatic building extraction from HRSI.
• The dual-branch architecture with DRA and LFE effectively captures multi-scale features.
• This research advances semantic segmentation techniques for remote sensing applications.