High-Performance Preprocessing of Architectural Drawings for Legend Metadata Extraction via OCR

Abstract

This paper describes the results of an investigation into methods of preprocessing architectural plots to enable them to be processed very quickly via OCR, detecting the region containing the relevant metadata legend and obtaining it in machine-readable form for e.g. automated folding and filenaming applications. We show how a processing pipeline adapted to this type of content can vastly decrease processing time, maintaining acceptable accuracy. Initial results show a reduction in total processing time from 2--3 minutes to around 15 seconds for most documents encountered, with the folding orientation being correctly detected in 78% of cases and the legend region being completely detected in 60% of cases, high enough for the use-case at hand.