Defect Image Generation for Deep Learning Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating training data for artificial intelligence algorithms to detect defects in products are limited in creating diverse defect images, leading to insufficient training and reduced accuracy in defect detection.
Innovation Solution
A method and system for generating defect images for training, which involves extracting a defect area from a sample image, transforming its shape, correcting it to match the target area, and synthesizing it with a base image to create diverse defect images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If common image augmentation methods (rotating, flipping, rescaling) are used to generate training data, then the generation process is simple and fast, but the diversity of defect images is limited and insufficient for充分 training
Solution Approach 1:
The patent segments the image processing into distinct components: defect area extraction from sample images, separate shape transformation operations, and synthesis with base images. This segmentation allows each component to be optimized independently, enabling both speed and diversity.
Solution Approach 2:
The patent introduces shape transformation as an additional dimension beyond traditional geometric augmentations. By transforming the shape of extracted defect areas while preserving their semantic content, the method creates fundamentally new defect patterns that differ from simple rotations or flips, significantly expanding defect image diversity.
2Manufacturing precision
If diverse defect images are generated through complex processing, then training data quality improves, but the generation time and computational complexity increase
Solution Approach 1:
The patent performs preliminary extraction of defect areas from sample images before synthesis. By pre-processing and storing defect areas separately, the actual synthesis process becomes faster as it only requires combining pre-extracted defect regions with base images, rather than performing complex processing during synthesis.
Solution Approach 2:
The patent copies extracted defect areas and applies shape transformations to create variations. This copying approach allows rapid generation of multiple defect instances from a single extracted defect region, maintaining high quality while reducing the need for extensive manual processing of each individual defect image.
3Ease of operation
If simple transformations are applied to defect images, then the processing is fast and simple, but the artificial intelligence algorithm cannot be sufficiently trained to detect various types and shapes of defects
Solution Approach 1:
The patent introduces dynamic shape transformation that can adaptively modify defect shapes while maintaining processing efficiency. The shape transformation operations can be applied with varying degrees and types of transformation, allowing the system to generate diverse defect patterns without requiring complex manual intervention for each transformation type.
Data Source
AI summary
A method of generating a defect image for deep learning and a system therefor are provided. The method and the system are intended to be used in generating training data for an artificial intelligence algorithm. More specifically, the training data are defect images required to train an algorithm that identifies a defect from a product.