Abstract
Graphics recognition remains a significant challenge compared to text-based OCR due to the high variability and domain-specific nature of technical drawings. This paper proposes a generalized architecture for "document mining" in the graphical domain. By combining hierarchical graph-based data models with a multi-layered processing pipeline, the system transforms hand-drawn architectural sketches into structured CAD data. We evaluate the platform using a custom database of architectural symbols, demonstrating near-perfect accuracy (99.08%) for on-line stroke-based recognition.
Introduction
While Optical Character Recognition (OCR) is a mature field, the interpretation of plans, maps, and engineering drawings—collectively known as Graphics Recognition—is often hindered by "brittle" systems that lack generalizability. A system optimized for musical scores, for instance, cannot easily process floor plans. Furthermore, the transition from paper to digital often excludes the user from the loop. This research addresses these gaps by proposing a general platform that supports both off-line (scanned) and on-line (digital pen) inputs, allowing for natural human interaction through a sketch-based interface.
System Architecture
The proposed platform is organized into a three-layer architecture designed to move from raw input to high-level semantic understanding.
- Acquisition: This layer handles input from diverse sources, including scanned paper documents and real-time digital pen/tablet data.
- Feature Extraction: A suite of specialized engines extracts four types of descriptors: vector-based (lines/arcs), perceptual grouping (parallelism/loops), pixel-based (Zernike moments/circular zoning), and dynamic (stroke sequence/speed).
- Knowledge Interpretation: This layer synthesizes descriptors into high-level entities using a hybrid approach, employing rule-based logic for structural elements and statistical classification for complex symbols.
Hierarchical Data Modeling
Central to the platform's ability to reason across different scales is its hierarchical graph model. Rather than treating a document as a flat collection of pixels or vectors, the system represents it as a graph where nodes and edges correspond to entities at varying levels of abstraction.
As illustrated in the data model, the system can represent simple points at one level and aggregate them into complex symbols or entire rooms at higher levels. This hierarchical structure mimics the human cognitive process of reading a technical plan, where individual lines are perceived as components of larger functional units.
Experimental Results
The system was evaluated using a custom ground-truth database of 4,200 symbol images across 25 architectural classes, drawn by 10 participants using a digital pen.
- Off-line Recognition: The system achieved a 69.3% accuracy rate for scanned images. This performance is attributed to high intra-class variability in hand-drawn styles; however, considering the "top two" candidates increased accuracy to 84%.
- On-line Recognition: By leveraging temporal stroke data, the system achieved a near-perfect recognition rate of 99.08%.
- Efficiency: The on-line engine processes approximately 200 images per second, supporting real-time interactive design.
Conclusion
The proposed platform demonstrates that a generalized, multi-layered architecture can effectively bridge the gap between messy hand-drawn sketches and structured digital models. The significant performance boost provided by temporal data highlights the value of on-line recognition in user-centric design tools. Future work will focus on personalizing models to individual drawing styles and expanding the descriptor library to support other graphical domains.
Summary updated