Chart Data Extraction Using CNN Decomposition and Regression Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current chart extraction technologies are limited in their ability to automatically and accurately interpret a wide range of graphic objects, such as pie and bar charts, especially those with variations in shape and position, and are not suitable for real-time industrial applications due to reliance on simplistic assumptions and lack of machine intelligence, leading to inefficient data capture and high processing times.
Innovation Solution
A machine learning-based method using a convolutional neural network that decomposes charts into sub-parts, performs optical character recognition with linear regression calibration, and aggregates data from images of pie and bar charts with tolerance to shape and position variations, enabling fast and fully automated data capture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If prior art algorithms use simplistic assumptions and arbitrary thresholds for chart extraction, then the algorithms are easier to implement, but they are restricted to ideal cases and cannot handle real-world chart variations
Solution Approach 1:
The patent transforms chart extraction from rule-based thresholding to a continuous learning process where parameters are automatically adjusted through training on diverse chart examples, enabling the system to adapt to various chart types and styles without manual parameter tuning
Solution Approach 2:
The patent replaces the mechanical rule-based extraction system with a machine learning model that learns patterns from data, substituting rigid algorithms with adaptive intelligent processing capable of handling real-world variations
2Adaptability or versatility
If prior art methods use deep learning models for high-level information extraction, then the system can handle diverse chart types, but the processing time is too long for real-time industrial applications
Solution Approach 1:
The patent segments the chart extraction task into distinct processing stages (chart detection, element identification, data extraction, value reading) and processes them in sequence, allowing optimization of each stage independently to reduce overall processing time while maintaining accuracy
Solution Approach 2:
The patent performs preliminary chart detection and classification before detailed data extraction, preparing the system in advance with metadata about chart structure and elements, which accelerates the subsequent extraction process by avoiding repeated analysis
3Measurement precision
If prior art algorithms assume high quality or vector format images, then text extraction accuracy is improved, but the system fails to handle real-world compressed or small size images
Solution Approach 1:
The patent incorporates image enhancement and preprocessing capabilities directly into the extraction system, allowing it to automatically improve the quality of input images through denoising, sharpening, and resolution adjustment before processing, making the system self-sufficient in handling various image qualities
4Measurement precision
If manual intervention is used to adjust extracted chart data, then extraction accuracy is improved, but the industrial benefit and time savings are drastically reduced
Solution Approach 1:
The patent implements automated validation and confidence scoring mechanisms that provide feedback on extraction quality, allowing the system to automatically correct or flag uncertain cases without human intervention, maintaining high accuracy while preserving processing efficiency
Data Source
AI summary
Using a convolutional neural network, a method for digitizing and extracting meaning from graphic objects such as bar and pie charts, decomposes a chart into its sub-parts (pie and slices or bars, axes and legends) with significant tolerance to the wide range of variations in shape and relative position of pies, bars, axes and legends. A linear regression calibration allows properly reading values even when there are many OCR failures.