PIDPipe: A Pipeline for Digitizing P&ID with Unsupervised Symbol Localization

Digitizing Piping and Instrumentation Diagrams (P&IDs) is challenging since the information is loosely structured in the document. Mostly supervised learning has been used by researchers to extract the information from P&IDs. However, having good quality labels, particularly, for localizing the symbols that require tight bounding boxes on the symbols, is difficult. In this research work, we propose a pipeline for digitizing P&IDs called PIDPipe with an unsupervised method for symbol localization. We analyze our method on a dataset of synthetic and real P&IDs. Our method achieved an accuracy of 71.3%, 72%, 52%, 74% and 81% for text detection, text recognition, symbol localization, and solid and broken line detection, respectively. This digitized information can be used towards the automated creation of Human Machine Interfaces and Digital Twins.