3D Texture Mapping for Pose-Normalized Industrial Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial imaging systems face challenges with part pose variation, image sensor pose variation, and background variation, which introduce unwanted variability into datasets, impacting the reliability of machine learning-based defect detection algorithms, particularly in high-volume manufacturing processes where rapid deployment of defect detection solutions is desired.
Innovation Solution
A method and system that estimates an optimal pose for a manufactured part by combining image data with a 3D model, maps texture data to polygonal facets, and renders virtual images for normalized views, thereby managing pose variability and eliminating unwanted data variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large training datasets are used to improve algorithm reliability, then defect detection reliability improves, but data production time and cost increase
Solution Approach 1:
The patent creates synthetic training images by rendering 3D models of parts with various defects under different lighting conditions, camera poses, and backgrounds. This copying approach generates unlimited training data without requiring actual defective parts, thereby improving algorithm reliability while avoiding the time and cost of collecting large datasets from production
Solution Approach 2:
The system pre-processes part geometry data and creates comprehensive 3D models with embedded defect information before training begins. By preparing the training data structure in advance through automated rendering pipelines, the system eliminates the need for time-consuming manual data collection and labeling during deployment
2Measurement precision
If complex machine learning algorithms are used to detect part defects, then detection accuracy improves, but computational processing time increases
Solution Approach 1:
The patent pre-computes and stores defect signatures and visual characteristics during the training phase using rendered images. This preliminary preparation of defect patterns allows the inspection algorithm to quickly match actual defects against pre-analyzed templates during real-time inspection, achieving high accuracy without excessive processing time
Solution Approach 2:
The inspection system divides the defect detection task into separate processing stages: feature extraction from images, comparison against trained models, and defect classification. This segmentation allows complex algorithms to process different aspects of defect detection in parallel, improving overall processing efficiency while maintaining accuracy
3Adaptability or versatility
If multiple imaging systems are used to capture comprehensive part information, then inspection coverage improves, but system complexity and cost increase
Solution Approach 1:
The patent implements a unified 3D modeling framework that can process and integrate data from multiple imaging modalities including optical cameras, laser scanners, and structured light systems. This universal approach allows different imaging systems to contribute to a single comprehensive part model, improving inspection coverage while avoiding the complexity of managing separate analysis pipelines for each system
Solution Approach 2:
The system merges multiple image datasets and 3D scan data into a single integrated textured 3D model of the part. By combining information from various imaging systems into one unified representation, the patent achieves comprehensive inspection coverage while simplifying the overall system architecture through centralized data fusion
Data Source
AI summary
System and method that includes: capturing an image of a manufactured part using an image sensing device, wherein an actual pose of the manufactured part relative to the image sensing device in the image can vary from an expected pose of the manufactured part relative to the image sensing device; estimating an optimal pose that represents a transformation of the expected pose to the actual pose based on (i) the image, (ii) initial pose data that indicates the expected pose, and (iii) a 3D model that models a model part that corresponds to the manufactured part as a set of polygonal facets; mapping texture data from the image to facets of the 3D model based on the optimal pose to generate a textured 3D model corresponding to the optimal pose; and rendering a virtual image for a further pose of the manufactured part based on the textured 3D model.