Related Experiment Video
Updated: Jul 16, 2026
08:27
Image Recognition and Parameter Analysis of Concrete Vibration State Based on Support Vector Machine
Published on: January 5, 2024
A novel semi-supervised strip steel surface defect classification model based on CycleGAN and ResNet-50
Liqiang Zhang1, Weilin Cao2, Youming Li2
1School of Artificial Intelligence, Neijiang Normal University, Neijiang, 641100, Sichuan, China. zhangxiaosuan_ai@163.com.
Scientific Reports
|July 10, 2026
Summary
This study introduces an improved semi-supervised learning model for steel strip surface defect classification. It enhances data diversity using CycleGAN and improves fine-grained feature recognition with ResNet-50, achieving high accuracy on the NEU-CLS dataset.
Area of Science:
- Materials Science
- Computer Science
- Artificial Intelligence
Background:
- Steel strip quality is vital for product reliability, necessitating accurate surface defect identification.
- Existing deep learning models struggle with limited, unlabeled industrial data and GAN-generated samples lacking diversity.
- Fine-grained texture similarities in defect data further degrade classification performance.
Purpose of the Study:
- To develop a novel semi-supervised learning model for enhanced steel strip surface defect classification.
- To address limitations in data diversity and fine-grained feature extraction in existing methods.
- To improve the generalization ability and accuracy of defect detection models in industrial settings.
Main Methods:
- A data augmentation strategy using Cycle Generative Adversarial Network (CycleGAN) to generate diverse, realistic unlabeled samples.
- Implementation of a ResNet-50 based network for robust feature extraction and classification of fine-grained defects.
- Training and evaluation using a combination of limited labeled and extensive unlabeled data from the NEU-CLS dataset.
Main Results:
- The proposed model achieved high performance metrics on the NEU-CLS dataset, including 92.91% accuracy, 89.64% precision, and 84.81% recall.
- The CycleGAN augmentation improved sample diversity and better simulated real industrial defect distributions.
- The ResNet-50 backbone effectively captured subtle, fine-grained features crucial for distinguishing similar defect types.
Conclusions:
- The developed semi-supervised model offers a competitive and effective solution for steel strip surface defect classification, particularly with limited labeled data.
- CycleGAN-based data augmentation and ResNet-50 feature extraction significantly enhance model performance and generalization.
- This approach demonstrates a promising direction for improving automated quality control in steel manufacturing.