Deep nested U-structure network with frequency attention for building semantic segmentation.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Automated building segmentation from remotely sensed imagery is crucial for various applications.
• Existing methods face challenges like incomplete extraction, inaccurate edges, and difficulty with irregular shapes.

Purpose of the Study:

• To develop a novel, end-to-end framework for enhanced building segmentation.
• To address limitations in accuracy, completeness, and prediction of irregular targets in current segmentation models.

Main Methods:

• Introduced a novel end-to-end residual U-structure within a U-Net architecture.
• Integrated a frequency attention module to focus on salient features and reduce noise.
• Employed a hybrid loss function to improve segmentation mask accuracy and completeness.

Main Results:

• The proposed framework demonstrated superior performance compared to baseline methods on four benchmark datasets.
• Achieved more complete internal extraction and higher accuracy in edge segmentation.
• Showcased improved prediction capabilities for irregular building targets.

Conclusions:

• The novel U-Net framework with residual U-structure, frequency attention, and hybrid loss significantly advances automated building segmentation.
• The approach offers a more robust and accurate solution for extracting building information from remote sensing data.