Automated Training Data Generation for Defect Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The process of creating training data for machine learning in inspection systems is cumbersome, requiring manual labeling of defects in images, which is time-consuming and labor-intensive, especially when dealing with large numbers of images.
Innovation Solution
A machine learning device and method that uses a marked symbol on the inspection target to efficiently create training data, where the symbol is detected and used to label images as either 'OK' or 'NG', allowing for automated generation of training images and labels, and optional removal or modification of the symbol in the image processing stage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling of defects in images is performed to create training data, then the accuracy of training data labeling is improved, but the time and labor required for creating training data increases significantly
Solution Approach 1:
The patent applies preliminary action by having inspectors mark defects on the actual inspection target before image capture. This pre-marking allows the marking information to be directly transferred to the image data, eliminating the need for subsequent manual labeling of images. The defect positions are predetermined and marked on the physical object, which then correlates directly with the captured image pixels, significantly reducing the time and labor required for training data preparation while maintaining labeling accuracy.
2Measurement precision
If inspectors view actual inspection targets and then label corresponding images, then the accuracy of defect identification is improved, but the workload becomes cumbersome due to needing to locate defect positions in both physical and image domains
Solution Approach 1:
The patent applies copying by transferring the marking information from the physical inspection target directly to the digital image domain. The markings made on the actual object are captured along with the image, creating a direct correspondence between physical defect locations and their digital representations. This eliminates the need for inspectors to manually locate and label defects in the image domain separately, as the marking information is already copied and associated with the correct image pixels.
3Measurement precision
If multiple images are captured and inspectors search through them to find defects, then the completeness of defect detection is improved, but the time and complexity of the labeling process increases proportionally
Solution Approach 1:
The patent applies preliminary action by marking all defects on the inspection target before capturing multiple images. This ensures that defect locations are predetermined and consistent across multiple image captures. When multiple images are taken from different angles or positions, the marking information serves as a reference that simplifies the labeling process, as inspectors only need to associate the pre-identified defect positions with the corresponding image data, rather than searching through multiple images to find all defects.
Data Source
AI summary
A machine learning device that creates training data to be used in machine learning includes: an image input unit that inputs an image which was obtained by capturing an inspection target on which a symbol indicating a defect is marked; and a creation unit that creates the training data based on the inputted image, in which the creation unit: creates training data consisting of a training image which is the image as inputted, and a label that retains a value of OK which signifies not having a defect, in a case of there not being the symbol in the image inputted; and creates training data consisting of a training image generated based on the image inputted, and a label that retains a value of NG signifying having a defect, in a case of there being the symbol in the image inputted.