Automated Digitization of Engineering Schematics

Abstract

Engineering schematics, such as electrical, mechanical, piping and instrumentation diagrams, record how industrial plants are built and operated, but most of them survive only as images or scanned sheets that software cannot read. Digitizing them by hand is slow and error-prone: an expert must find and classify hundreds of symbols, read dense technical text, and work out which label belongs to which component. Progress with learning-based methods has been held back on two fronts at once. There are almost no annotations that connect a text label to its symbol, and the drawings themselves are usually confidential, so even unlabeled sheets rarely reach the public domain. We address this with a system that turns a drawing into a structured, queryable graph: it detects and classifies the graphical components with an object detector, recovers the technical text, and then resolves which label belongs to which component. Our contributions are threefold: (i) the first at-scale dataset of manually annotated text-to-symbol links for industrial schematics; (ii) a complete, deployable digitization system combining tiled detection with sliced inference, off-the-shelf OCR, and a text-to-symbol association stage; and (iii) a rigorous, leakage-free benchmark of association methods. Under an observable-only candidate protocol, we find that on logic circuits association is dominated by geometry: a simple pairwise model reaches about 99% top-1 and a graph neural network matches but does not exceed it, whereas the denser P&IDs still benefit from a geometric rule-based chain. Detection reaches an mAP@50 of 0.995 on logic circuits and about 0.91 across the 107-class P&ID taxonomy. The system produces a partial semantic graph; connecting lines and flow direction are not extracted.