Document Data Extraction with Optical Character Recognition

The modern solutions for fetching content from PDF documents or images includes fetching document’s multiple features like processing using Natural Language Processing (NLP), border-less table contents, processing it’s stacked lines, fetching document layout information and working on region-based features from documents. There are new challenges like fetching a table’s content where the table is included in borders of other content. In this paper, we have used LayoutLM for parsing the content and RetinaNet for table detection in a particular document. The pretrained models were trained on very less data-set to decide for large scale use. We also have discussed architecture of LayoutLM and tesseract that we used in building the tool with discussing output of some other approaches.