AI-powered text extraction from engineering drawings

Abstract

Engineering systems contain vast amounts of data in the form of engineering documents, which contain textual data about the systems they represent. In cases where engineering documents are transformed from one format to another, this textual data might be lost. For a system such as eShare that utilizes the textual data from engineering drawings to create links between the model and documents, losing this textual data causes issues. This thesis investigates the previous implementations that are being used for extracting textual content from engineering drawings via a literature review. Each proposed implementation is benchmarked with a test set that mimics eShare’s use-case. This dataset contains various types of engineering documents in differing sizes and formats. Results from the literature review are used for a case-study in which a plan is proposed for eShare to create an OCR solution that will extract textual content from engineering drawings that do not have it included in the metadata. The literature review was able to conclude that the latest solution from PaddleOCR, PP-OCRv5 is the optimal solution to be used for eShare. However, there were limitations found with the ready-to-use pipeline, which was to be addressed. The proposed pipeline utilizes the detection and recognition model from PP-OCRv5 but implements custom intermediate processing to handle word-level detection, which is required by eShare. The proposed solution was able to extend F1 accuracy with 8% to reach an accuracy of 89% compared to PP-OCRv5’s 81%.