A pavement crack synthesis method based on conditional generative adversarial networks.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Limited datasets hinder effective training of machine learning models for asphalt pavement crack detection.
• Existing methods struggle with generating specific types of pavement cracks, impacting research diversity.

Purpose of the Study:

• To develop a Generative Adversarial Network (GAN) capable of generating realistic asphalt pavement crack images.
• To augment existing road crack datasets for improved crack detection and classification research.
• To selectively generate images for specific crack categories like transverse and longitudinal cracks.

Main Methods:

• Leveraged five open-source road crack datasets to construct a comprehensive image dataset.
• Employed a Conditional Generative Adversarial Network (CGAN) incorporating fully connected, convolutional, and attention mechanisms.
• Implemented a Self-Attention Generative Adversarial Network (SAGAN) by integrating self-attention mechanisms with CGAN.

Main Results:

• The SAGAN model effectively generated high-quality, realistic pavement crack images.
• The model successfully addressed data limitations and enabled selective image generation based on category.
• Optimized GAN parameters and hyperparameters improved the realism of generated images.

Conclusions:

• SAGAN offers a robust solution for generating diverse and category-specific pavement crack images.
• The integration of self-attention and conditional mechanisms significantly enhances generative model performance for road infrastructure analysis.
• This approach facilitates advancements in automated road inspection and maintenance through improved dataset availability.