Engineering Drawing Text Extraction Using Horizontal Edge Projections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR systems face challenges in accurately extracting text from engineering drawings, particularly those with varying fonts and sizes, and struggle with noise levels and minimal distance between text and graphical components, leading to inaccurate outputs.
Innovation Solution
A method and system that classify components in engineering drawings as textual or non-textual, using deep learning classifiers and horizontal edge projections to identify word elements, which are then processed to enhance OCR accuracy by reducing noise and isolating text regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If engineering drawings are directly passed to the OCR system without removing graphics or drawings, then the processing speed is maintained, but the OCR accuracy deteriorates due to noise from graphical elements
Solution Approach 1:
The patent segments the engineering drawing into multiple components by classifying each component as textual or non-textual using deep learning classifiers. This segmentation allows the system to separate text regions from graphical elements, improving OCR accuracy by processing only relevant text components while filtering out noise from drawings and annotations.
Solution Approach 2:
The patent extracts textual components from the engineering drawing by identifying and isolating text regions using horizontal edge projections and overlap detection. This extraction process removes graphical elements and non-textual components before passing the cleaned text regions to the OCR system, thereby improving accuracy without requiring complex full-drawing processing.
2Adaptability or versatility
If traditional text extraction methods are used, then the system is simpler, but it fails to handle varying fonts and sizes robustly
Solution Approach 1:
The patent uses deep learning classifiers that can adapt to varying parameters such as font types, sizes, and styles by training on diverse text samples. The system dynamically adjusts its classification parameters to recognize text components regardless of their specific visual characteristics, enabling robust handling of varying fonts and sizes while maintaining reliability in noisy environments.
Solution Approach 2:
The patent performs preliminary classification of components as textual or non-textual before OCR processing. This preliminary action using deep learning models prepares the data by identifying text regions in advance, allowing the system to handle varying fonts and sizes more effectively by focusing computational resources on text detection rather than processing all graphical elements.
3Measurement precision
If distance-based text region identification is used, then the method is straightforward, but it provides inaccurate output when distance between text components and drawing components is minimal or zero
Solution Approach 1:
The patent introduces horizontal edge projections as an additional dimensional approach to text region identification. Instead of relying solely on distance metrics in the spatial domain, the system projects text components horizontally and identifies overlaps in the projected dimension. This dimensional transformation enables accurate text region identification even when text components and drawing components are adjacent or overlapping in the original spatial arrangement.
Data Source
AI summary
Examples of present disclosure relate to method and system to extract text from engineering drawing for performing accurate OCR. Initially, for the extraction, image of engineering drawing is received with a plurality of components. Each of the plurality of components in the image is classified to be one of a textual component and a non-textual component. At least one word element for textual components from the plurality of components is identified based on segmentation of the plurality of components. The segmentation is performed by drawing a plurality of horizontal edge projections of a predefined length for each of the textual components. Further, the textual components is identified to be associated with the at least one word element when horizontal edge projection of each of the textual components overlaps with adjacent textual component. The at least one word element is provided as extracted text for performing OCR on the engineering drawing.