Surface Defect Inspection Using Unsupervised Image Reconstruction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing material surface inspection techniques require laborious human labeling of datasets or struggle with engineering effective algorithms, leading to increased production costs and limitations in defect detection.

Innovation Solution

An automated defect inspection mechanism using an unsupervised neural network that trains itself to generate defect-free or reduced-defect images, allowing for defect detection without labeled training data, leveraging generative adversarial networks and convolutional denoising neural networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If human labeling is used to create training datasets, then defect detection accuracy is improved, but labor cost and time consumption increase significantly

Engineering Contradiction:
Improvedefect detection accuracyVSAvoidlabeling time consumption
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system uses unsupervised learning where the neural network automatically learns defect patterns from unlabeled images without requiring human annotators. The model self-trains by identifying anomalies through reconstruction error, eliminating the need for manual labeling while maintaining detection accuracy.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent uses generative models to create synthetic defect-free images that serve as training data. By generating realistic defect-free samples through GANs or diffusion models, the system obtains sufficient training data without manual labeling, resolving the contradiction between accuracy and labeling effort.

Inventive Principle:
Principle #26Copying

2Productivity

If hand-crafted algorithms are used for defect detection, then implementation speed is improved, but detection effectiveness deteriorates due to engineering difficulty

Engineering Contradiction:
Improveimplementation speedVSAvoiddetection effectiveness
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent replaces traditional mechanical image processing algorithms with deep learning-based unsupervised models. The neural network automatically learns complex defect patterns from data, substituting hand-crafted feature engineering with automated feature learning, thereby improving both effectiveness and maintaining implementation speed through standardized training pipelines.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Measurement precision

If labeled training data is required, then model accuracy is improved, but data preparation complexity increases

Engineering Contradiction:
Improvemodel accuracyVSAvoiddata preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system employs unsupervised learning algorithms that automatically learn from unlabeled images. The neural network performs self-training by comparing reconstructed images with original inputs, identifying defects through reconstruction errors without requiring any manual labeling or complex data annotation processes.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent generates synthetic defect-free images using generative models to create training datasets. By copying and transforming real defect-free images through GANs or diffusion models, the system obtains sufficient diverse training data without the complexity of manual data collection and labeling.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3739513B1Surface defect identification method and apparatus
Publication Date: 2026.02.18 FUJITSU LTD
  • EP3739513B1 patent drawingFigure 1
  • EP3739513B1 patent drawingFigure 2~3D
  • EP3739513B1 patent drawingFigure 4A~5

AI summary

Embodiments include a method of processing images of material surfaces to identify defects on the imaged material surface, the method comprising automatically training a neural network to generate reduced-defect versions of input training images of material surfaces; acquiring an image of a subject material surface; inputting the acquired image to the neural network to generate a reduced-defect version of the acquired image; comparing the generated reduced-defect version of the acquired image with the acquired image to identify differences; identifying defects on the subject material surface at locations of the identified differences.