Appearance Inspection Image Generation for Defect Training Diversity

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing pseudo-defective image creation devices generate a limited number of pseudo-defective images, leading to insufficient accuracy in identifying defects due to the scarcity of training images of defective products.

Innovation Solution

An image generation device that generates training images by extracting features from non-defective products and combining them with non-defective images to create pseudo-defective images, enhancing the variety and accuracy of training data through hue replacement, noise addition, boundary combination, and background integration.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If pseudo-defective images are created from a small number of defective product images, then the number of training images is increased, but the types of training images remain limited and identification accuracy is not sufficiently improved

Engineering Contradiction:
Improvenumber of training imagesVSAvoidtypes of training images
Core Design Contradiction:
Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The patent creates synthetic defective images by copying and combining features from non-defective product images with defect information. Instead of merely duplicating existing defective images, the system generates new pseudo-defective images by transferring defect characteristics onto non-defective product backgrounds, thereby increasing both quantity and diversity of training data

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent applies parameter changes by modifying various image characteristics including color, brightness, contrast, and defect position parameters when generating pseudo-defective images. This allows the creation of diverse training images with varying visual parameters while maintaining realistic product appearances, directly addressing the limitation of uniform pseudo-defective image types

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If the number of defective products is small, then the number of training images obtained by imaging defective products is insufficient, but increasing the number of defective products is not feasible

Engineering Contradiction:
Improvenumber of training imagesVSAvoiddefective products
Core Design Contradiction:
Quantity of substanceVSLoss of substance

Solution Approach 1:

The patent uses image copying and synthesis techniques to create artificial defective product images without requiring physical defective products. By copying defect patterns and superimposing them onto images of non-defective products, the system generates sufficient training data while avoiding the need to accumulate rare defective physical samples

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent introduces computer-generated defect masks and transfer images as intermediaries between non-defective product images and synthetic defective product images. These intermediary elements enable the transformation of available non-defective images into training data for defect detection without requiring actual defective products

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3675033B1Image generation device and appearance inspection device
Publication Date: 2026.05.20 TDK CORP
  • EP3675033B1 patent drawingFigure 1
  • EP3675033B1 patent drawingFigure 2~3
  • EP3675033B1 patent drawingFigure 4~5

AI summary

An image generation device and an appearance inspection device which can increase the number of types of training image are provided. An image generation device includes an image acquisition unit configured to acquire a first image; and an image generator configured to generate a training image based on the first image, wherein the first image depicts a first defect included in a first product of a different type the product to be inspected, and the training image is a image to be read by an identifier that identifies whether an inspection image obtained by imaging the product to be inspected depicts a same type of defect as the first defect and to be used by the identifier to perform learning.