Frequency Domain Defect Inspection Training Data Creation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing defect inspection methods for equipment like piping require numerous sample images to ensure accuracy, which is challenging due to low occurrence frequency of certain defects and confidentiality issues, leading to inefficiencies in defect detection.
Innovation Solution
A training data creation method that involves frequency distribution analysis and band-pass filtering to standardize image shades, allowing for accurate defect inspection even with limited sample images, by selecting a frequency bandwidth signal and creating training data based on defect information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If many sample images are collected to ensure inspection accuracy, then defect detection accuracy is improved, but the time and resources required for data collection increase significantly
Solution Approach 1:
The patent applies preliminary action by performing frequency distribution analysis and band-pass filtering on training images before they are used for machine learning training. This preprocessing standardizes the shade characteristics of defect images in advance, allowing the system to achieve high detection accuracy with fewer sample images. The frequency components are analyzed and filtered beforehand to emphasize defect-related features while suppressing irrelevant variations, thus reducing the need for extensive data collection.
2Measurement precision
If many sample images with various shades are collected to cover defect variations, then inspection accuracy is improved, but the difficulty of obtaining sufficient samples increases due to low defect occurrence frequency
Solution Approach 1:
The patent applies parameter changes by transforming the training images through frequency domain analysis and band-pass filtering. This changes the parameter representation of the images from spatial domain to frequency domain, allowing the system to extract and emphasize defect-related frequency components. By adjusting the band-pass filter parameters, the system can adapt to different defect types and conditions, effectively handling the variability in defect shades and appearances without requiring numerous physical samples.
3Device complexity
If traditional defect inspection methods are used without frequency analysis, then the inspection process is simpler, but the accuracy of defect detection decreases when sample images are limited
Solution Approach 1:
The patent applies segmentation by dividing the image processing into distinct frequency components through Fourier transform and band-pass filtering. Instead of processing the entire image as a single unit, the method segments the frequency spectrum to isolate and emphasize defect-related components. This segmentation allows the machine learning model to focus on specific frequency ranges that are most indicative of defects, improving detection accuracy without requiring excessive complexity in the overall inspection process.
Data Source
AI summary
Provided are a training data creation method and device, and a defect inspection method and device capable of securing the accuracy of defect inspection even though the number of samples of a defect to be used in creating training data is small.The training data creation method includes acquiring a training-use image including a received light image created based on reflected light or transmitted light from an inspection object having a defect obtained by irradiating the inspection object with light rays or radiation, executing frequency distribution analysis on the training-use image, receiving an input of a parameter for designating a frequency bandwidth, selecting a frequency bandwidth signal from an analysis result of the frequency distribution analysis according to the frequency bandwidth designated by the parameter, acquiring defect information indicating a defect for an image corresponding to the frequency bandwidth signal, and creating training data to be used in learning of a defect inspection device, which inspects a defect of the inspection object, based on the defect information.