With the growth of information flow through supply chains, we address the issue of semantically segmenting tabular data from document flows. This challenge is abstract in definition as there is no guideline to what defines a table, therefore we primarily applied deep learning methods to compare performances, as well as a morphology-based method to compare several methods for the task of table structure detection. Taking advantage of transfer learning, we found that Mask-RCNN was the most capable network at abstracting to segmentation of tables. Due to the large variation in sizes between columns and rows, most networks failed to detect both with equal effectiveness. However with Mask-RCNN 97.01% precision and 98.28% recall were attained, which put it far ahead of other models in row detection, cementing Mask-RCNN as the most effective choice in this task.
Paper
Full text
Deep Learning for Recognizing the Anatomy of Tables on Datasheets
Semantic Scholar · Computer Science · 2019
Abstract
With the growth of information flow through supply chains, we address the issue of semantically segmenting tabular data from document flows. This challenge is abstract in definition as there is no guideline to what defines a table, therefore we primarily applied deep learning methods to compare performances, as well as a morphology-based method to compare several methods for the task of table structure detection. Taking advantage of transfer learning, we found that Mask-RCNN was the most capable network at abstracting to segmentation of tables. Due to the large variation in sizes between columns and rows, most networks failed to detect both with equal effectiveness. However with Mask-RCNN 97.01% precision and 98.28% recall were attained, which put it far ahead of other models in row detection, cementing Mask-RCNN as the most effective choice in this task.