Improved Pixel-Level Pavement-Defect Segmentation Using a Deep Autoencoder.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Vision-based systems excel in quality inspection, surpassing human speed and accuracy.
• Infrastructure maintenance, like road upkeep, is labor-intensive and time-consuming.
• Automated pavement-quality evaluation is crucial for efficient infrastructure management.

Purpose of the Study:

• To develop an automated system for pavement-quality evaluation using pixelwise defect segmentation.
• To enhance the U-Net deep autoencoder architecture for improved defect extraction.
• To evaluate the performance of the proposed model against standard U-Net.

Main Methods:

• Utilized a U-Net deep autoencoder for pixelwise pavement defect segmentation.
• Incorporated residual connections, atrous spatial pyramid pooling (with parallel and "Waterfall" connections), and attention gates into the U-Net architecture.
• Conducted experiments on CrackForest, Crack500, GAPs384, and mixed datasets for validation.

Main Results:

• The enhanced neural network configurations demonstrated improved segmentation performance compared to the original U-Net.
• The proposed model achieved better defect extraction capabilities.
• Performance improvements were achieved with no significant increase in computational overhead.

Conclusions:

• The modified U-Net architecture effectively improves pavement defect segmentation accuracy.
• The integration of advanced components enhances defect extraction without substantial computational cost.
• This approach offers a promising solution for efficient and accurate pavement-quality evaluation.