Image Generation System for Machine Learning Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image generation systems struggle to produce images suitable for machine learning, particularly in generating diverse and accurate learning data for models like those used in weld appearance testing.
Innovation Solution
The system includes a first image acquirer, a second image acquirer, and an image processor that performs image transformation and superposition processing. The image transformation processing involves dividing an extracted part of the first image into segments and altering parameters such as grayscale, location, size, and orientation of these segments. The superposition processing then combines these transformed segments with the second image to generate a superposed image suitable for machine learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If simple image replacement is used to generate new images, then image generation is easy and quick, but the generated images are not suitable for machine learning
Solution Approach 1:
The patent divides the extracted part of the first image into multiple segments, allowing independent transformation of each segment. This segmentation enables complex transformations (rotation, scaling, grayscale changes) that create diverse training data suitable for machine learning, while still building upon the simple replacement foundation.
Solution Approach 2:
The patent transforms image segments by changing multiple parameters including rotation angle, scale factor, grayscale values, and position. These parameter transformations generate varied versions of the same defect patterns, creating diverse training data that improves machine learning model robustness while maintaining the efficiency of automated generation.
2Adaptability or versatility
If diverse transformations are applied to image segments, then learning data diversity is improved, but processing complexity increases
Solution Approach 1:
By dividing the extracted part into segments, the patent applies transformations to manageable units rather than the entire image. This reduces computational complexity compared to transforming whole images multiple times, while still achieving diverse training data through segment-level variations.
Solution Approach 2:
The patent applies transformations to only the extracted defect portions rather than entire images. This partial action approach generates sufficient diversity for machine learning training without the excessive computational cost of transforming complete images with multiple defects and backgrounds.
3Adaptability or versatility
If extracted parts are divided into multiple segments, then transformation flexibility is improved, but processing time increases
Solution Approach 1:
The patent segments extracted defect parts into multiple smaller units that can be transformed independently and in parallel. This segmentation enables flexible transformations of individual defect characteristics while reducing overall processing time through parallel computation of segment transformations.
Data Source
AI summary
An image generation system includes a first image acquirer, a second image acquirer, and an image processor. The image processor performs image transformation processing and superposition processing. The image transformation processing includes generating, based on a first image, a transformed image by subjecting a predetermined extracted part of a first object to image processing. The image transformation processing includes: processing of dividing the extracted part into a plurality of segments; and processing of changing at least one parameter selected from the group consisting of a grayscale, a location, a size, and an orientation of at least one segment belonging to the plurality of segments.