TableBorderNet: A Table Border Extraction Network Considering Topological Regularity.

Area of Science:

• Computer Vision
• Digital Image Processing
• Machine Learning

Background:

• Accurate table border extraction from scanned road engineering drawings is vital for digital transformation and intelligent infrastructure.
• Degraded borders, blur, and adjoining characters pose significant challenges for automated table parsing.
• Existing methods struggle with visual distortions and require extensive annotated data, limiting real-world applicability.

Purpose of the Study:

• To develop a robust framework for precise table border extraction from complex and degraded engineering drawings.
• To improve the accuracy and reliability of automated table structure delineation in engineering archives.
• To reduce the reliance on manually annotated datasets for training deep learning models.

Main Methods:

• Proposed TableBorderNet, a semantic segmentation framework guiding feature extraction along explicit row and column directions.
• Introduced a topology-aware loss function to ensure structural consistency in degraded inputs.
• Implemented a generative self-supervised strategy to simulate degradation patterns and reduce annotation needs.

Main Results:

• Achieved an Intersection-over-Union (IoU) of 94.2% for table border extraction.
• Demonstrated a low topological error of 1.07%, indicating high structural accuracy.
• Outperformed existing approaches in precision and robustness on complex visual conditions.

Conclusions:

• TableBorderNet offers a practical and scalable solution for precise table border extraction in engineering drawings.
• The framework significantly advances the digital transformation of engineering archives and supports data-driven road asset management.
• The method's ability to handle degraded inputs and minimize annotation requirements enhances its real-world utility.