Advancing Table Detection In Engineering Drawings Through Synthetic Data Generation

The digitization of engineering drawings via Table OCR has been pivotal in extracting tabular data, crucial for structured information in assembly diagrams. However, the intricate layout styles and varied table structures within these drawings present a challenge for accurate table detection. Existing methods lack adaptability and generality, compelling the need for a novel approach. This paper introduces a solution leveraging synthetic engineering drawing generation to create a diverse dataset. A material library, comprising figures, text, and tables extracted from genuine drawings, forms the basis. The synthetic drawing generation involves randomizing tables with varying structures and integrating them into a wrap-around layout. Using YOLO object detection models trained on this synthetic dataset, the study evaluates different training strategies. Results indicate the effectiveness of synthetic data in enhancing table detection, especially when combined with real data for model fine-tuning.