Image Segmentation for Synthetic Defect Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data generators for machine learning face challenges in improving the accuracy of learning data, especially when the amount of defective data is small or when there is a low degree of similarity between original and synthesized images.
Innovation Solution
A data creation system that includes a first image acquirer, a second image acquirer, a segmenter, a range generator, and a creator, which acquires images, segments them into regions, generates range patterns, and superposes the particular part on the second image to create learning data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a data generator locates defect spots from a statistical point of view, then it can handle cases with small amounts of defective data, but the accuracy of the generated learning data decreases
Solution Approach 1:
The patent segments the image into multiple regions and performs defect detection on each region separately. This allows the system to handle diverse defect locations and types more effectively, improving the accuracy of learning data generation even when defective data is scarce. The segmentation enables region-specific processing that captures local characteristics better than statistical methods alone.
Solution Approach 2:
The patent introduces an intermediary process that generates candidate defect spots based on region segmentation results, then validates these candidates using multiple criteria including background pattern similarity. This intermediary validation step improves accuracy by filtering out statistically generated but unrealistic defect locations.
2Reliability
If the data generator searches for spots with similar background patterns, then it can improve the realism of synthesized images, but the complexity of the image processing increases
Solution Approach 1:
The patent divides the image into regions first, then performs background pattern comparison only within corresponding regions. This segmented approach reduces the complexity of image processing by limiting the comparison scope, while still maintaining the realism of synthesized images through region-specific pattern matching.
Solution Approach 2:
The patent applies different processing strategies to different regions based on their characteristics. For regions where background pattern similarity is important, it performs detailed pattern comparison. For other regions, it uses simpler methods. This local quality approach maintains realism where needed while controlling overall processing complexity.
3Adaptability or versatility
If the data generator synthesizes images with low similarity to original images, then it can increase data diversity, but the accuracy of locating synthesized spots decreases
Solution Approach 1:
The patent performs preliminary region segmentation and candidate spot identification before synthesizing diverse images. By establishing reference regions and candidate locations in advance, the system maintains accuracy in locating synthesized spots even when generating highly diverse synthetic data with low similarity to originals.
Solution Approach 2:
The patent uses copying of region structures and background patterns from original images to create synthetic images. By copying the fundamental structural elements while varying surface details, the system achieves data diversity while maintaining accurate correspondence between original and synthesized spots through the preserved region structure.
Data Source
AI summary
A data creation system includes a first image acquirer, a second image acquirer, a segmenter, a range generator, and a creator. The first image acquirer acquires a first image representing a first object including a particular part. The second image acquirer acquires a second image representing a second object. The segmenter divides at least one of the first image or the second image into a plurality of regions. The range generator generates, based on a result of segmentation obtained by the segmenter, a single or plurality of range patterns. The creator superposes, in accordance with at least one range pattern belonging to the single or plurality of range patterns, the particular part on the second image to create a single or plurality of superposed images and output the single or plurality of superposed images as learning data.