Table Extraction with Table Data Using VGG-19 Deep Learning Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing methods for tabular data processing struggle with diverse table layouts, styles, and noise.
• Task-specific features and model architectures limit accurate table structure extraction.

Purpose of the Study:

• To develop a comprehensive deep learning methodology for precise row and column extraction from document images containing tables.
• To improve table structure recognition and address limitations in existing datasets.

Main Methods:

• A deep learning model combining table detection, structure recognition, and semantic rule-based row extraction.
• Utilized transfer learning with VGG-19 for model fine-tuning.
• Enhanced the Marmot dataset with additional table structure annotations, including column detection.

Main Results:

• Achieved state-of-the-art performance on the Marmot dataset for table structure extraction.
• Demonstrated the effectiveness of the proposed deep learning methodology.
• Successfully expanded the Marmot dataset's annotation scope.

Conclusions:

• The proposed deep learning approach offers a robust solution for accurate table row and column extraction from document images.
• The enhanced Marmot dataset provides a valuable resource for future research in table understanding.
• Transfer learning further boosts the model's performance in table structure recognition.