Generative Template Defect Detection for Part Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems for defect detection in manufactured parts face challenges in accurately identifying defects due to insufficient training data and poor training performance, leading to reduced inspection accuracy.
Innovation Solution
A part inspection system utilizing a generative neural network architecture that generates a template image from 'good' images, allowing for defect detection by comparing input images to the template, thereby identifying defects without requiring images of defective parts for training, and overlaying defect identifiers on the output image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image processing systems are trained using both good and bad images, then the system can learn to identify defects, but the training process becomes time-consuming and requires sufficient numbers of both good and bad images which are often insufficient
Solution Approach 1:
The patent extracts and removes the requirement for bad images from the training process. Instead of training on both good and bad images, the system only uses good images to create a template, then detects defects by comparing new images against this template. This eliminates the time-consuming collection and processing of bad images while maintaining defect detection capability.
Solution Approach 2:
The system performs preliminary action by creating a template image from good images before actual defect detection begins. This pre-established template serves as a reference for all subsequent inspections, eliminating the need to retrain or reprocess bad images for each inspection task.
2Reliability
If traditional systems gather many images including both good and bad images for training, then the training data becomes sufficient, but the data collection and processing becomes complex and resource-intensive
Solution Approach 1:
The patent extracts the essential training requirement down to only good images, removing the complexity of managing and processing both good and bad images. The system achieves reliable training by focusing solely on creating an accurate template from good images, simplifying the entire training workflow.
Solution Approach 2:
Instead of the traditional approach of training by showing both good and bad examples, the patent inverts the methodology by training only on good examples and using deviation from this norm to detect defects. This inversion simplifies data requirements and processing complexity.
3Productivity
If the algorithm uses traditional defect detection methods, then it can process images, but the accuracy performs poorly when training data is insufficient
Solution Approach 1:
The patent creates an ideal copy or template of a defect-free part from training images. This template copy serves as the perfect reference against which all inspection images are compared, enabling accurate defect detection without requiring actual defective images for training.
Data Source
AI summary
A part inspection system includes a vision device configured to image a part being inspected and generate a digital image of the part. The system includes a part inspection module communicatively coupled to the vision device and receives the digital image of the part as an input image. The part inspection module includes a defect detection model. The defect detection model includes a template image. The defect detection model compares the input image to the template image to identify defects. The defect detection model generates an output image. The defect detection model configured to overlay defect identifiers on the output image at the identified defect locations, if any.