Ray-Traced Defect Image Generation for Label-Free ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating training data for supervised machine learning in defect classification require significant labor, are prone to inconsistent labeling, and struggle to collect data for low-frequency defects under varying conditions, making them inefficient and costly.
Innovation Solution
Generate training image data through ray tracing simulations that model defect and inspection target surface models, incorporating lighting and imaging systems, to produce pseudo images without manual labeling, allowing for flexible simulation of defect shapes and conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If manual labeling of large amounts of images is performed to obtain sufficient training data, then the quantity of training data is improved, but the labor cost and time consumption increase significantly
Solution Approach 1:
The patent uses optical simulation to generate synthetic training images that copy the characteristics of real defect images without requiring manual labeling. The simulation creates virtual images of defects under various lighting and inspection conditions, providing sufficient training data automatically.
Solution Approach 2:
The system performs preliminary action by pre-collecting defect information and creating defect models in advance. These models are then used to generate training images through simulation, eliminating the need for time-consuming manual labeling of large image datasets.
2Measurement precision
If manual labeling is performed to ensure accurate defect classification, then the labeling accuracy is improved, but the inconsistency between different labelers increases
Solution Approach 1:
The simulation system generates training images with ground truth labels automatically, ensuring perfect consistency. The synthetic images replicate real defect characteristics while providing standardized labels without human intervention, eliminating inter-labeler variability.
3Quantity of substance
If real defect images are collected for low-frequency defects, then the representativeness of training data is improved, but the data collection time and cost increase
Solution Approach 1:
The system creates synthetic images of rare defects through simulation, copying the visual characteristics of low-frequency defects without needing to collect actual examples. This provides sufficient training data for rare defect types immediately, regardless of their actual occurrence frequency.
Solution Approach 2:
The system performs preliminary action by pre-defining defect models that include rare defect types. These models can generate training images for low-frequency defects on demand, eliminating the need for lengthy data collection periods.
4Adaptability or versatility
If multiple images are collected for different inspection target characteristics and capturing conditions, then the coverage of training scenarios is improved, but the complexity of data collection and processing increases
Solution Approach 1:
The optical simulation system serves multiple functions simultaneously: it generates images for different defect types, lighting conditions, inspection angles, and target characteristics through a single unified platform. This eliminates the need for separate data collection efforts for each scenario.
Solution Approach 2:
The system performs preliminary action by pre-configuring defect models and inspection parameters. Once set up, the simulation can generate training images for various conditions automatically without requiring complex data collection procedures for each scenario.
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
Reduces the cost and effort of generating training data by producing accurate, labeled-free pseudo images that cover various defect types and conditions, ensuring sufficient data for rare defects.
Implementation Method 1
a plurality of light rays are traced, and illuminance in each of the pixels is calculated based on an intensity and the number of light rays entering the pixels
Data Source
Figure 1~2B
Figure 3A~3C
Figure 4A~4D
AI summary
A generation system (1) for generating training image data for supervised machine learning for training a defect classifier to be applied to visual inspection arranges a defect model arbitrarily selected from a storage (7) to store defect models obtained by modeling shapes and optical characteristics of defects and an inspection target surface model arbitrarily selected from a storage (7) to store inspection target surface models obtained by modeling shapes and optical characteristics of inspection target surfaces in any space in which ray tracing is performed by ray tracing simulation software in which a lighting optical system, an imaging optical system (60), and an imaging sensor (63) including a plurality of pixels are modeled. A plurality of light rays are traced, and illuminance in each of the pixels is calculated based on an intensity and the number of light rays entering each of the pixels. Data generating means (8) is provided for generating, based on the illuminance, pseudo image data as training image data.