TableBorderNet: A Table Border Extraction Network Considering Topological Regularity

Accurate extraction of table borders in scanned road engineering drawings is crucial for the digital transformation of engineering archives, which is an essential step in the development of intelligent infrastructure systems. However, challenges such as degraded borders, image blur, and character adjoining often hinder the precise delineation of table structures, making automated parsing difficult. Existing solutions, including traditional OCR tools and deep learning methods, struggle to consistently delineate table borders in the presence of these visual distortions and fail to perform well without extensive annotated datasets, which limits their effectiveness in real-world applications. We propose TableBorderNet, a semantic segmentation framework designed for precise border extraction under complex visual conditions. The framework captures structural context by guiding convolutional feature extraction along explicit row and column directions, enabling more accurate delineation of table borders. To ensure topological consistency in complex or degraded inputs, a topology-aware loss function is introduced, which explicitly penalizes structural discontinuities during training. Additionally, a generative self-supervised strategy simulates common degradation patterns, allowing the model to achieve strong performance with minimal reliance on manually annotated data. Experiments demonstrate that the method achieves an Intersection-over-Union of 94.2% and a topological error of 1.07%, outperforming existing approaches. These results underscore its practicality and scalability for accelerating the digitization of engineering drawings in support of data-driven road asset management.