Text Detection and Post-OCR Correction in Engineering Documents

Abstract

As the amount of born-analog engineering documents is still very large, the information they contain can not be processed by a machine or any automatic process. To overcome this, a whole process of digital transformation must be implemented on this type of documents. In this paper, we propose to detect and recognize all textual entities present on this type of documents. They can be part of technical details about a technical diagrams, bill of material or functional descriptions, or simple tags written in a standardized format. These texts are present in the document in an unstructured way, so that they can be located anywhere on the plan. They can also be of any size and orientation. We propose here a study allowing the text detection and recognition with or without associated semantics (symbolic annotations and dictionary words). A solution coupling a text detector based on a deep learning architecture, an open-source OCR for string recognition and an OCR post-correction process based on text clustering is proposed as a first step in the digital transformation process of industrial plans and P&ID schemes. The results applied to a database of 30 images of industrial maps and plans from different industries (oil, gas, water...) are very promising and close to 84% of correct detection and 82% of correct tags (and lexicon-free words) recognition after post-correction.