CascadeTabNet: An approach for end to end table detection and structure recognition from image-based documents

An automatic table recognition method for interpretation of tabular data in\ndocument images majorly involves solving two problems of table detection and\ntable structure recognition. The prior work involved solving both problems\nindependently using two separate approaches. More recent works signify the use\nof deep learning-based solutions while also attempting to design an end to end\nsolution. In this paper, we present an improved deep learning-based end to end\napproach for solving both problems of table detection and structure recognition\nusing a single Convolution Neural Network (CNN) model. We propose\nCascadeTabNet: a Cascade mask Region-based CNN High-Resolution Network (Cascade\nmask R-CNN HRNet) based model that detects the regions of tables and recognizes\nthe structural body cells from the detected tables at the same time. We\nevaluate our results on ICDAR 2013, ICDAR 2019 and TableBank public datasets.\nWe achieved 3rd rank in ICDAR 2019 post-competition results for table detection\nwhile attaining the best accuracy results for the ICDAR 2013 and TableBank\ndataset. We also attain the highest accuracy results on the ICDAR 2019 table\nstructure recognition dataset. Additionally, we demonstrate effective transfer\nlearning and image augmentation techniques that enable CNNs to achieve very\naccurate table detection results. Code and dataset has been made available at:\nhttps://github.com/DevashishPrasad/CascadeTabNet\n