Research on High-Fidelity Industrial Defect Sample Generation Method Based on Crop-Integrate Generative Adversarial Network
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
In industrial product quality inspection, the automation of surface defect identification is limited by the scarcity of defect samples and insufficient diversity in defect morphology. Traditional data augmentation methods struggle to generate diversified samples that conform to the real defect distribution, while existing Generative Adversarial Network (GAN) methods focus more on global image generation, leading to lower fidelity of local defect features and unnatural background integration. To address this issue, this study proposes a Crop-Integrate Generative Adversarial Network (CIGAN) based on improved StyleGAN2, featuring innovative techniques such as cropping based on local defect blocks, dual-discriminator cooperative optimization, and a dynamic seam-sealing mechanism based on Poisson blending, achieving automated synthesis of high-fidelity industrial defect samples. Experiments on the NEU steel surface defect dataset demonstrate that the network has improved image similarity and defect diversity. Sensitivity analysis further indicates that each mechanism is not simply stacked but a systematically tailored solution around the characteristics of industrial image defects. CIGAN provides a high-fidelity data augmentation solution for low-sample industrial scenarios, advancing intelligent manufacturing quality inspection systems towards low-cost and high robustness.
Fields of Study & Subjects
Industrial Vision Systems and Defect Detection
Generative Adversarial Networks and Image Synthesis
Infrastructure Maintenance and Monitoring
Publication details
- Journal / source
- IEEE International Conference on Power, Electronics and Green Energy (ICPEGE)
- Pages
- 321–330
- Publication date
- 28 January 2026
- Language
- EN