Structured Light 3D Product Capture for Synthetic Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The development and implementation of machine learning-based quality control algorithms for product inspection is hindered by the need for extensive annotated image data, class imbalances, and the inability to adapt to changing conditions, leading to high costs and delays.
Innovation Solution
A method involving dual optical captures of real objects using structured projection patterns and varying illumination to generate synthetic data that accurately represents three-dimensional geometry and surface properties, enabling the creation of realistic virtual scenes for training machine learning algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If extensive real-world image data is collected and annotated for training machine learning algorithms, then the training data quantity is sufficient, but the time and cost for data acquisition and annotation increase significantly
Solution Approach 1:
The patent creates synthetic images by rendering virtual copies of products with defects in a controlled virtual environment, eliminating the need to capture and annotate real defective products. The synthesis unit generates training images by combining product models with defect models, providing sufficient training data without time-consuming real-world data collection
Solution Approach 2:
The system pre-defines defect models and product models in advance, allowing synthetic training data to be generated on-demand without requiring actual defective products to be captured and annotated. This preliminary preparation of virtual models enables rapid generation of annotated training data with any required defect types
2Reliability
If real-world data is used for training, then the data represents actual product variations, but class imbalances occur when certain defect types are underrepresented
Solution Approach 1:
The synthesis unit allows independent control of defect parameters including type, size, position, and quantity. This enables deliberate creation of balanced datasets by adjusting parameters to ensure equal representation of all defect classes, overcoming the inherent class imbalances in real-world captured data
3Extent of automation
If machine learning algorithms are trained with fixed datasets, then the training process is complete and deterministic, but the algorithms cannot adapt to changing production conditions or new defect types
Solution Approach 1:
The system enables dynamic generation of training data by allowing users to modify defect models and product models after initial training. When new defect types or production conditions arise, new synthetic images can be generated on-demand by updating the virtual models, allowing the system to adapt without complete retraining
Solution Approach 2:
The synthesis unit can autonomously generate additional training data or fine-tune existing models by incorporating new defect models or product variations without requiring external data collection or expert annotation, enabling the system to self-adapt to changing conditions
4Measurement precision
If extensive expert knowledge is required for data annotation and algorithm development, then high-quality training data can be created, but the complexity and cost of implementation increase
Solution Approach 1:
By using automated rendering of virtual product copies with programmatically applied defects, the system eliminates manual annotation entirely. Defect characteristics such as position, size, and type are precisely controlled through software parameters rather than requiring expert annotators to manually label real images
Solution Approach 2:
The patent replaces the manual mechanical process of expert annotation with automated computer-based synthesis. The synthesis unit uses software algorithms to automatically generate defect images with precise control over all annotation attributes, substituting human expert labor with automated computational processes
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach reduces the need for extensive real-world data acquisition and annotation, improves algorithm adaptability, and enhances the accuracy and flexibility of machine learning models for quality control.
Implementation Method 1
projecting a structured projection pattern, preferably a fringe projection, preferably several different structured projection patterns, preferably fringe projections, successively onto the object
Implementation Method 2
illuminate the object from different spatial directions by means of a plurality of light sources
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
The present invention relates to a method for generating synthetic data of objects (2), preferably of products (2), comprising at least the following steps: • Positioning (000) of an object (2), • Projecting (100) a structured projection pattern, preferably a fringe projection, preferably several different structured projection patterns, preferably fringe projections, successively onto the object (2), • preferably, first optical acquisition (120) of the object (2) together with the structured projection pattern, • Determining (140) the three-dimensional geometry of the object (2) based on the optically acquired structured projection pattern, • with the object (2) in an unchanged position, a second optical acquisition (220) of the object (2) from the same perspective as in the first optical acquisition (120).• Determining (240) the surface properties of the object (2) based on the optically acquired image data from the second optical acquisition (220), • Assigning (300) the determined three-dimensional geometry of the object (2) and the determined surface properties of the object (2), and • Generating (500) synthetic data of the object (2) based on the assigned data.