Beyond the Legend

Abstract

The digitization of industrial diagrams has seen increasing attention across many engineering domains due to these documents often being the backbone for downstream applications such as maintenance, revision work and validation. But working with these drawings can often times be expensive, time-consuming and repetitive. While related work has explored automated symbol detection in domains such as piping and instrumentation diagrams (P&IDs), electrical housing diagrams are less studied even though they have many of the same challenges. This thesis investigates symbol localization and reference-guided classification in electrical housing diagrams, focusing on the use of diagram legends as references. The work compares two methodologies, a more traditional template matching approach and a two-stage approach, where symbol regions are detected, by a generic symbol region of interest detector, and then classified using legend-based references. The two-stage method is tested using both template matching and Siamese-networkbased classification. By relying on legend symbols, the proposed methods reduces the dependence on annotated training datasets reflecting realistic scenarios where data, more often than not, is very limited. The generic regions of interest detector was able to localize a large portion of the relevant symbols, reaching a maximum recall of 95.42%. Classification on the proposed regions showed promising performance, with results comparable to the traditional template matching approach. The reference-guided two-stage method also offers a more flexible structure by separating localization from classification as well as incorporating deep-learning models, providing stronger potential for scalability, speed, refinement and versatility.