Table Recognition in Scanned Documents

Invoices are so vastly used in business. For each invoice, an employee has to verify carefully written data including date, legal, and the courtesy amount present in each table. However, this task is not only time-consuming but also prone to inaccuracies and errors, especially when it comes to processing a massive amount of invoices. A smart capture system is required to facilitate processing invoices automatically and it is more challenging since relevant data are not narrative but arranged in tables. Although it is true that OCR (Optical Character Recognition) is able to read and capture data, it suffers from inefficiency in table locating and loses structural features of tabular data. Table recognition is widely carried out using deep learning and heuristics and a better result was reached as humans would. In this paper, we present a part of a smart capture system for invoices which is based on table recognition workflow for scanned invoices. This workflow consists of three main steps: the first step is a prepossessing step which is used to enhance the quality of scanned invoices. The second step is a deep learning-based table detection approach where we use DocCutout and DocCutmix for data augmentation. The third step is a heuristic-based table structure recognition approach. The presented approaches are evaluated on public data sets.