Table Detection for Improving Accessibility of Digital Documents using a Deep Learning Approach

Assistive technologies play an important role in improving the quality of life of people with disabilities. In this work, we developed a system for the retrieval of table information from digital documents for use in screen readers used by visually impaired people. The proposed methodology takes advantage of computer vision techniques with a deep learning approach to make documents accessible instead of the classical rule-based programming approach. We explained in detail the methodology that we used and how to objectively evaluate the approach through entropy, information gain, and purity metrics. The results show that our proposed methodology can be used to reduce the uncertainty experienced by visually impaired people when listening to the contents of tables in digital documents through screen readers. Our table information retrieval system presents two improvements compared with traditional approaches of tagging text-based portable document format (PDF) files. First, our approach does not require supervision by sighted people. Second, our system is capable of working with image-based as well as text-based PDFs.