Automatic Crack Detection on Road Pavements Using Encoder-Decoder Architecture.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Road pavement integrity is crucial for safety and durability, with cracks being an early indicator of failure.
• Conventional crack detection methods are often slow, costly, and struggle with complex crack features like low contrast and poor continuity.
• Automated crack detection is essential for efficient pavement monitoring and maintenance.

Purpose of the Study:

• To develop an advanced automated crack detection algorithm for pavement surfaces.
• To address the limitations of traditional methods in identifying complex crack patterns.
• To leverage deep learning for improved accuracy and efficiency in pavement distress analysis.

Main Methods:

• Proposed a novel encoder-decoder architecture named U-Hierarchical Dilated Network (U-HDN).
• Incorporated hierarchical feature learning and dilated convolution for multi-contextual crack feature extraction.
• Utilized a multi-dilation module to integrate crack features of various sizes.

Main Results:

• The U-HDN method demonstrated high performance in pixel-wise crack detection on public datasets.
• Achieved superior results compared to existing methods on the same image datasets.
• Effectively extracted and fused multi-scale and multi-contextual features for accurate crack identification.

Conclusions:

• The U-HDN algorithm offers a robust and efficient solution for automated pavement crack detection.
• The proposed architecture successfully handles complex crack characteristics.
• This deep learning approach enhances pavement monitoring capabilities and contributes to road safety.