Introduction
In the modern landscape of Industry 4.0, manufacturing processes increasingly rely on digital twins and Computer-Aided Tolerance (CAT) software. However, a significant hurdle remains: millions of "legacy" mechanical engineering drawings (MEDs) that exist only as paper blueprints or static image files. Manually transcribing the complex Geometrical Tolerance Specification Callouts (GTSCs) from these documents into digital systems is a labor-intensive bottleneck, prone to errors that can necessitate intervention in up to 40% of cases.
The Challenge of Automated Extraction
Automating the recognition of GTSCs is technically demanding. These callouts are not simple text; they are structured blocks containing a mix of specialized symbols, numbers, and letters. They are often arranged in multi-line formats, frequently tilted due to scanning or photography angles, and often obscured by noise or overlapping leader lines from the drawing itself.
Proposed Methodology: A Hybrid Pipeline
To address these challenges, the researchers developed a multi-stage pipeline that integrates traditional digital image processing with advanced deep learning. The workflow is designed to handle the structural complexity of engineering documents while maintaining high recognition accuracy.
1. Pre-processing and Localization
The initial stage focuses on cleaning the input image and isolating the tolerance blocks. Gaussian filtering and binarization are used to eliminate noise. The system then employs morphological operations and Hough line detection to identify the rectangular frames of the GTSCs amidst the clutter of a full drawing.
2. Tilt Correction and Segmentation
For drawings captured via mobile devices, the model applies perspective transformation formulas to correct any tilt. Once straightened, the system performs advanced segmentation:
- Multi-line Splitting: Hough transforms identify and remove horizontal separator lines within multi-line blocks.
- The "Drop-Fall" Algorithm: To solve the problem of "touching characters"—where ink bleeds together or characters touch the box borders—the researchers implemented a heuristic that simulates a water drop falling through gaps to find the optimal segmentation path.
3. Recognition via Inception-AlexNet
The core of the recognition engine is Inception-AlexNet, a modified version of the classic AlexNet architecture. By replacing standard convolutional layers with Inception modules, the model can perform parallel convolutions of different sizes. This allows it to capture multi-scale features, which is critical for distinguishing between similar-looking technical symbols (such as different types of run-out or orientation tolerances). The addition of Batch Normalization (BN) and a reduced fully connected layer (2048 neurons) helps prevent overfitting and improves generalization.
Experimental Results
The model was trained and tested on a dataset of over 20,000 augmented images covering 106 categories of symbols, numbers, and letters.
- Recognition Accuracy: The Inception-AlexNet achieved a 98.3% accuracy rate for individual character recognition, outperforming the standard AlexNet by 1.8%.
- System-Wide Performance: When evaluated on the entire pipeline—from localization through to semantic recognition—the system achieved a success rate of 96.73% on real-world technical documents.
- Efficiency: The architecture demonstrated faster convergence and lower training loss, suggesting its suitability for industrial deployment.
Conclusion and Future Work
This research provides a robust solution for digitizing legacy engineering data, bridging the gap between paper-based documentation and modern automated manufacturing. While the system handles complex interferences and leader lines effectively, future work will focus on further "lightening" the model for real-time use on mobile devices and integrating it directly into automated online measurement systems.
Summary updated