TableStrRec: framework for table structure recognition in data sheet images

Billions of documents in data sheet format are shared between various organizations across the globe on a daily basis. The essential information in these documents is presented in tabular format. Extracting and assimilating this information can help organizations make data-driven decisions. Solutions for detecting tables in document images have been well explored. Thus, in this work, we propose TableStrRec, a deep learning-based approach to recognize the structure of such detected tables by detecting rows and columns. TableStrRec comprises two Cascade R-CNN architectures, each with a deformable backbone and Complete IOU loss to improve their detection performance. One architecture detects and classifies rows as regular rows (rows without a merged cell) and irregular rows (groups of regular rows that share a merged cell). The second architecture detects and classifies columns as regular columns (columns without a merged cell) and irregular columns (groups of regular columns that share a merged cell). Both architectures work in parallel to provide the results in a single inference. We show that utilizing TableStrRec to detect four classes of objects improves the table structure recognition performance on three public test sets. We achieve 90.5 % and 89.6 % weighted average F1 scores on the ICDAR2013 test set for rows and columns, respectively. On the TabStructDB test set, we achieve 72.7 % and 78.5 % weighted average F1 score for rows and columns, respectively. We also evaluate the proposed method under the FinTabNet dataset using the structure-only TEDS score, achieving 98.34%, which can outperform most state-of-the-art benchmark models.