Related Experiment Video
Updated: Jun 22, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
515
DG2GAN: improving defect recognition performance with generated defect image sample
Fuqin Deng1,2, Jialong Luo1, Lanhui Fu1
1School of Mechanical and Automation Engineering, The Wuyi University, Jiangmen, 529000, China.
Scientific Reports
|June 26, 2024
Summary
This study introduces DG2GAN, a novel method for generating diverse, high-quality defect images to address data imbalance in manufacturing. This approach significantly enhances deep-learning-based surface defect recognition accuracy and precision.
Area of Science:
- Computer Vision
- Machine Learning
- Materials Science
Background:
- Surface defect recognition in manufacturing faces challenges with imbalanced, low-quality, and non-diverse image data.
- Existing data augmentation methods often fail to generate realistic and varied defect images.
Purpose of the Study:
- To develop an advanced defect generation method to improve deep-learning-based surface defect recognition.
- To overcome limitations of data imbalance, insufficient diversity, and poor quality in defect image datasets.
Main Methods:
- A novel Generative Adversarial Network (GAN) named DG2GAN is proposed, incorporating cycle consistency loss, DJS optimized discriminator loss, and a new DG2 adversarial loss.
- DG2GAN generates defect images from defect-free images, ensuring data diversity and high quality.
Main Results:
- DG2GAN produced defect images with superior quality and diversity compared to existing methods.
- Data augmentation using DG2GAN on CrackForest and MVTec datasets improved defect recognition accuracy from 86.9% to 94.6% and precision from 59.8% to 80.2%.
Conclusions:
- The DG2GAN method effectively generates high-quality, diverse synthetic defect images.
- Employing DG2GAN for data augmentation significantly enhances the performance of surface defect recognition systems in manufacturing.
