Virtual Defect Image Generation for X-ray Inspection Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing X-ray inspection devices face challenges in creating a large amount of training data necessary for machine learning, particularly in distinguishing between foreign matter and cavities inside articles, which often lack sharp edges in X-ray inspection images.
Innovation Solution
An image generation device that acquires a first image of an article without defects and generates a second image with a virtual defect, altering pixel values to simulate different X-ray attenuation rates within the article, thereby creating training data without physically introducing defects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If physical samples with defects are created to generate training data, then the training data can be obtained, but it is not easy to create a large number of such samples and the process is time-consuming
Solution Approach 1:
The patent uses image processing to create virtual copies of defect-containing images by digitally manipulating pixel values in regions of interest from original inspection images. Instead of physically creating defect samples, the system generates synthetic training data by copying and modifying existing image data, thereby easily producing large amounts of training data without the difficulty of physical sample preparation
Solution Approach 2:
The system performs preliminary image processing and defect simulation before actual machine learning training. By pre-generating virtual defect images with controlled parameters (defect type, size, position, X-ray attenuation characteristics), the system prepares comprehensive training data in advance, eliminating the need for time-consuming physical sample creation during the training process
2Quantity of substance
If physical samples with defects are created to generate training data, then the training data can be obtained, but the process requires significant time to create large amounts of data
Solution Approach 1:
The system rapidly generates training data by digitally copying and modifying pixel values in image regions, replacing the time-consuming physical sample preparation process. This digital copying approach allows instantaneous generation of multiple defect variations from single original images
Solution Approach 2:
The system efficiently generates diverse training data by changing parameters such as defect size, position, shape, and X-ray attenuation coefficients through digital image processing. This parameter-based approach allows rapid generation of numerous training samples without physical manipulation, significantly reducing data preparation time
3Measurement precision
If image processing parameters are adjusted to detect foreign matter and cavities, then inspection accuracy can be improved, but it is difficult to set parameters because these defects do not appear with sharp edges
Solution Approach 1:
The patent introduces a learned model (AI) as an intermediary between the raw X-ray images and the inspection decision. This learned model automatically learns optimal detection parameters and features from training data, eliminating the need for manual parameter setting. The intermediary handles the complexity of detecting subtle defects without sharp edges, allowing standard image processing parameters to be used while achieving high inspection accuracy
Solution Approach 2:
The system transforms the inspection approach by changing from fixed parameter-based image processing to adaptive parameter learning through machine learning. The learned model automatically adjusts detection parameters based on the specific characteristics of foreign matter and cavities, making the inspection system adaptable to various defect types without manual parameter tuning
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 enables the easy creation of a large amount of training data, improving the accuracy of machine learning models used in X-ray inspection devices for detecting foreign matter and cavities.
Implementation Method 1
an X-ray irradiation unit that irradiates the article conveyed by the conveying unit with X-rays, an X-ray detection unit that detects X-rays transmitted through the article
Implementation Method 2
an image generation unit configured to change at least part of pixel values of a region corresponding to the article in the first image acquired by the image acquisition unit, and generate a second image containing a virtual defect in the article. The image generation unit generates the second image in which the virtual defect is contained inside the article and an attenuation rate of the X-rays in the virtual defect is different from an attenuation rate of the X-rays in the article
Data Source
AI summary
A server includes an image acquisition unit configured to acquire a non-defective article image on the basis of X-rays transmitted through an article not containing a defect, and an image generation unit configured to change at least part of pixel values of a region corresponding to the article in the non-defective article image acquired by the image acquisition unit and generate a defective article image in which the article contains the virtual defect. The image generation unit generates the defective article image in which the virtual defect is contained inside the article and an attenuation rate of the X-rays in the virtual defect is different from an attenuation rate of the X-rays in the article.