What is it about?

In large-scale construction projects, such as building power plants or industrial facilities, engineers must manage thousands of technical drawings. A critical but tedious task involves analyzing legend sheets, which explain the symbols and components used in the blueprints. A major challenge is that these legend sheets do not follow a universal standard. Every project uses different templates, styles, and formats. Furthermore, they are often visually confusing because they place different sets of information side-by-side without clear borders or lines to separate them. Traditional AI and computer programs often struggle with these inconsistent and messy layouts, leading to errors that require hours of manual correction. Our research introduces ICICLE, an intelligent system designed to automate this process. Instead of using traditional programming or expensive AI training for every new project, we developed a method called ICMAP. This technique allows an engineer to take just one example drawing from a specific project, highlight the important areas with simple digital markers like colored circles, and provide a brief instruction. The AI then learns from this single example how to correctly sort and extract information from all other documents in that project, easily adapting to whatever unique format is being used. We tested ICICLE on real-world data from four distinct industrial projects. Our system achieved between 96 percent and 100 percent accuracy, significantly outperforming standard AI methods and traditional text-recognition tools. By turning a manual process into an automated task that takes minutes, this work helps the engineering industry save significant time and costs. It serves as a practical blueprint for how modern AI can be used to solve complex, real-world problems without needing specialized technical expertise or constant reprogramming for new formats.

Featured Image

Read the Original

This page is a summary of: ICICLE: An Interactive VLM-based System for Information Extraction from Ambiguous Engineering Diagram Legends, March 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3748522.3779736.
You can read the full text:

Read

Contributors

The following have contributed to this page