3D Defect Model Generation for Piston Visual Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional visual inspection methods generate defect images based on two-dimensional shapes, leading to inaccuracies when inspecting surfaces with unevenness and varying illumination, requiring extensive data preparation and collection of defective product samples.
Innovation Solution
A visual inspection apparatus using machine learning with three-dimensionally pre-generated defect models to generate two-dimensional defect images and simulate defective product sample images, reducing the need for actual sample collection and improving inspection accuracy by considering luminance distributions and angles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If two-dimensional defect images are used for inspection, then the inspection process is simple, but the inspection accuracy deteriorates on surfaces with unevenness and varying illumination
Solution Approach 1:
The patent transitions from two-dimensional defect images to three-dimensional defect models. By pre-generating defect models in 3D space that incorporate surface unevenness and illumination characteristics, the system achieves higher inspection accuracy while maintaining process simplicity through automated model generation and machine learning.
2Measurement precision
If three-dimensionally pre-generated defect models are used, then the inspection accuracy improves, but the data preparation complexity increases
Solution Approach 1:
The patent performs preliminary generation of three-dimensional defect models before the actual inspection process. By pre-generating comprehensive defect models that account for various surface conditions and illumination scenarios, the system prepares all necessary data in advance, eliminating the need for extensive real-time data collection and reducing on-site preparation complexity.
Solution Approach 2:
The patent creates virtual copies of defects through three-dimensional modeling instead of requiring physical defective samples. These synthesized defect models replicate real defect characteristics while allowing flexible generation of diverse training data without needing to collect extensive physical samples from production lines.
3Measurement precision
If extensive defective product samples are collected for machine learning, then the model accuracy improves, but the time and effort for sample collection increases
Solution Approach 1:
The patent replaces physical sample collection with virtual defect model generation. By synthesizing defective product images from three-dimensional defect models combined with surface images, the system generates unlimited training data without needing to collect actual defective samples from production lines, dramatically reducing time and effort.
Solution Approach 2:
The patent performs preliminary generation of diverse defect variations through three-dimensional modeling and image synthesis. By pre-generating a comprehensive set of training images covering various defect types, positions, and lighting conditions, the system prepares all necessary training data before the machine learning process, eliminating time-consuming sample collection during deployment.
4Adaptability or versatility
If the inspection system is adapted to new products or equipment, then the versatility improves, but the re-learning process requires extensive sample collection again
Solution Approach 1:
The patent creates a universal three-dimensional defect model framework that can be applied across different products and equipment. By modeling defects in 3D space rather than product-specific 2D images, the system can generate training data for new products by combining the same defect models with new surface images, enabling rapid adaptation without re-collecting physical samples.
Data Source
AI summary
A visual inspection apparatus inspects a surface of a piston based on a captured image acquired by a camera and a learning result acquired by conducting machine learning using a plurality of defective product sample images, which is each generated by combining a two-dimensional image of a defect image that is generated based on a three-dimensionally pre-generated defect model with an image of the surface of the piston.