Chart De-Rendering AI Training With Synthetic Ground Truth Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for constructing data sets for chart de-rendering AI models suffer from limited data diversity, lack of meta information, and difficulty in categorizing charts by style, which hinders accurate evaluation and training.
Innovation Solution
A system and method that generates data sets including line and meta information as ground truth, allowing for diverse chart styles and accurate classification, using a data set generation model to output chart images with selected parameters, and evaluates AI model performance through a performance evaluation model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data is collected by crawling a specific website, then data collection is simple, but the amount of data is limited and diversity is lacking
Solution Approach 1:
The patent uses a chart rendering engine to generate synthetic chart images that copy the visual appearance of real charts while allowing controlled variation of parameters. This enables generation of large quantities of diverse training data without crawling websites, resolving the contradiction between data quantity and diversity.
Solution Approach 2:
The system varies parameters such as chart types, color schemes, layouts, and data values to generate diverse chart images from the same underlying data structure. This parameter variation approach enables unlimited data diversity while maintaining consistent data quality and metadata completeness.
2Loss of information
If data is collected by crawling, then collection process is straightforward, but meta information is often not properly included
Solution Approach 1:
The system pre-generates complete metadata including chart titles, axis labels, legends, and data values during the synthetic chart generation process, before the charts are used for training. This preliminary inclusion of all metadata ensures completeness without requiring complex post-processing or crawling operations.
Solution Approach 2:
The chart rendering engine automatically generates both the visual chart image and its corresponding metadata structure in a self-contained process, eliminating the need for external crawling operations. The system serves itself by producing complete training data pairs with full metadata annotation.
3Measurement precision
If chart images of similar style are collected for comparison, then accurate evaluation is possible, but conventional crawling methods make this difficult
Solution Approach 1:
The system controls chart style parameters (color schemes, line styles, font types, layouts) to generate sets of charts with similar visual characteristics. By systematically varying only specific parameters while holding others constant, the system creates controlled experimental groups for accurate model evaluation without complex crawling strategies.
Solution Approach 2:
The chart rendering engine serves multiple functions: generating training data, creating evaluation datasets with controlled styles, and producing test charts. This universal generation capability enables accurate comparative evaluation while simplifying the data collection process into a single unified system.
Data Source
Figure 1~2
Figure 3
Figure 4A~4B
AI summary
A system, method, and program for evaluating the performance of an artificial intelligence model that de-renders a chart or for training the artificial intelligence model by constructing a data set including chart information is disclosed. The system includes a memory in which a data set generation model and an AI model are included, and a processor configured to execute or train the AI model and execute a performance evaluation model, wherein the data set generation model stores line information, which is information about at least one line of a chart, and meta information, which is information about meta data, as ground truth (GT), stores an image formed using the GT as a chart image, and outputs the GT and the chart image as a data set, and the AI model receives the chart image stored in the data set as input and outputs a data format in which information of the chart is predicted. The parameters included in the line information may include an X-axis value of a chart line, a function, a coefficient of the function, a color or a shape of a line or a point, etc. and the parameters included in the meta information may include a chart title, an X-axis name, a Y-axis name, a legend, etc. Furthermore, the data format output from the AI model may be used to evaluate the performance of the AI model or to train the AI model.