GAN-Based Pseudo Defect Image Generation for AI Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost of manufacturing numerous test pieces required to develop screening AI for defect detection from inspection images is a significant challenge, as it increases development costs.
Innovation Solution
An image generation device and method that uses an image generation algorithm trained on defect images with labels to create pseudo defect images on background images, allowing for efficient training of screening AI without the need for physical test pieces.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a large number of test pieces with artificial defects are manufactured to train screening AI, then the training data quantity is sufficient, but the development cost increases
Solution Approach 1:
The patent uses generative adversarial networks (GANs) to create synthetic defect images that copy the characteristics of real defect images. The generator network learns from real defect images and generates synthetic versions that preserve defect features while eliminating the need for physical test piece manufacturing. This copying approach provides sufficient training data quantity without incurring manufacturing costs
Solution Approach 2:
The patent replaces the mechanical process of physically manufacturing test pieces with artificial defects with a computational process using GANs. Instead of mechanically creating defect samples through material processing, the system uses neural networks to generate synthetic defect images, substituting physical manufacturing with information processing to reduce development costs
2Reliability
If physical test pieces are manufactured for AI training, then real defect samples are obtained, but the manufacturing process becomes complex
Solution Approach 1:
The GAN-based system creates accurate copies of real defect characteristics in synthetic images. The generator network learns the distribution and features of real defects from training data and reproduces them synthetically, maintaining defect quality and reliability without requiring complex physical manufacturing processes to create test pieces
3Adaptability or versatility
If numerous test pieces are produced for screening AI development, then comprehensive defect detection capability is achieved, but the development time increases
Solution Approach 1:
The patent replaces time-consuming physical test piece manufacturing and defect creation processes with rapid computational generation using GANs. The system can generate diverse defect samples including various types, shapes, and positions synthetically in parallel, achieving comprehensive defect detection capability without the sequential time constraints of physical manufacturing
Solution Approach 2:
The GAN model is trained once on a relatively small set of real defect images to learn defect characteristics. After this preliminary training phase, the generator can rapidly produce unlimited synthetic defect variations without requiring additional physical manufacturing time. This preliminary action of model training enables fast subsequent generation of diverse defect samples for comprehensive AI training
Data Source
AI summary
This image generation device comprises an image generation unit for using an image generation algorithm (AR)—which has been trained on the basis of defect images (PD) that are inspection images showing defects (D) and label images (PL) obtained by adding, to the defect images, labels (Lr) corresponding to the types and shapes of the defects-to input a label image for image generation that has been created through the addition of a desired label to a background image and generate a pseudo-defect image by drawing, on the background image, a pseudo defect corresponding to the label added to the label image for image generation. The image generation unit draws a type of pseudo defect that corresponds to the color of the label.