Related Experiment Video
Updated: May 7, 2025
09:52
Generation of Shear Adhesion Map Using SynVivo Synthetic Microvascular Networks
Published on: May 25, 2014
8.8K
Efficient dataset extension using generative networks for assessing degree of coating degradation around scribe
Dominik Stursa1, Pavel Rozsival1, Petr Dolezel1
1Faculty of Electrical Engineering and Informatics, University of Pardubice, Pardubice, Czechia.
Frontiers in Artificial Intelligence
|December 30, 2024
Summary
This study introduces a new method using generative adversarial networks (GANs) to create synthetic data for training AI models to detect surface degradation on coil coatings. Carefully managed synthetic data significantly improves AI model accuracy and reduces manual annotation efforts.
Area of Science:
- Materials Science
- Computer Science
- Artificial Intelligence
Background:
- Semantic segmentation of surface degradation in industrial applications like coil coating is crucial for quality control.
- Acquiring large, annotated datasets for training deep learning models is a significant challenge due to manual annotation costs and time.
- Existing datasets may lack diversity, limiting the generalization capabilities of trained models.
Purpose of the Study:
- To present a novel methodology for dataset augmentation in semantic segmentation of coil-coated surface degradation.
- To investigate the effectiveness of Deep Convolutional Generative Adversarial Networks (DCGAN) in generating realistic synthetic data for this task.
- To evaluate the impact of synthetic data augmentation on the performance of state-of-the-art segmentation models (U-net, DeepLabV3).
Main Methods:
- Dataset augmentation using DCGAN to generate synthetic input-target pairs for coil-coated surface degradation.
- Training and evaluating U-net and DeepLabV3 models on both original and augmented datasets.
- Systematic experimentation to determine optimal ratios of synthetic to real data for performance improvement.
- Exploration of different GAN architectures for their suitability in generating diverse and realistic samples.
Main Results:
- The introduction of synthetic data generated by DCGAN significantly improves the performance of U-net and DeepLabV3 models in detecting surface degradation.
- Optimal performance gains in accuracy and F1-score were observed when the ratio of synthetic to real data ranged between 0.2 and 0.5.
- The study identified advantages and limitations of various GAN architectures regarding the realism and diversity of generated samples.
- The proposed augmentation method demonstrated scalability for industrial applications.
Conclusions:
- Dataset augmentation using DCGAN provides a scalable and effective solution to enhance semantic segmentation models for coil-coated surface degradation assessment.
- The findings reduce the burden of manual data annotation while improving model generalization and segmentation accuracy.
- This approach enables more efficient and accurate degradation detection, benefiting industries reliant on coil coatings.
