Latent Diffusion Models to Enhance the Performance of Visual Defect Segmentation Networks in Steel Surface...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Real-world defect data for industrial components is often scarce and imbalanced, hindering the training of robust deep learning models.
• Synthetic data generation presents a viable solution to augment limited datasets and improve model generalization.

Purpose of the Study:

• To investigate the efficacy of state-of-the-art latent diffusion models, specifically stable diffusion, for generating synthetic defect images.
• To enhance the robustness of visual defect segmentation in manufacturing components by utilizing AI-generated data.

Main Methods:

• Fine-tuned stable diffusion using the LoRA (Low-Rank Adaptation) technique on the NEU-seg dataset.
• Integrated synthetic images generated by stable diffusion into training datasets for DeepLabV3+ and FPN segmentation models at varying ratios.
• Evaluated the impact of synthetic data augmentation on model performance using mean Intersection over Union (mIoU).

Main Results:

• Augmenting training datasets with synthetic images generated by stable diffusion led to significant improvements in mIoU for both DeepLabV3+ and FPN models.
• The proposed approach achieved notable gains in defect segmentation accuracy, with improvements of 5.95% and 6.85% in mIoU over models trained solely on real data.

Conclusions:

• Latent diffusion models, such as stable diffusion, are effective tools for generating high-quality synthetic data to improve industrial defect detection.
• Synthetic data augmentation enhances the diversity and quantity of training data, leading to more accurate and reliable visual defect segmentation in manufacturing settings.