Detection of Highway Pavement Damage Based on a CNN Using Grayscale and HOG Features

Area of Science:

• Civil Engineering
• Computer Vision
• Machine Learning

Background:

• Deep learning methods, particularly Convolutional Neural Networks (CNNs), are used for highway pavement damage detection.
• Current CNN methods using raw image data demand high-performance hardware and significant processing time.
• Simplifying CNN structures with preprocessed data is crucial for application in common scenarios.

Purpose of the Study:

• To develop a rapid and efficient highway pavement damage detection method.
• To reduce the computational demands of deep learning models for pavement damage detection.
• To enable pavement damage detection on common hardware configurations.

Main Methods:

• Proposed a novel detection method combining CNNs with grayscale and Histogram of Oriented Gradients (HOG) features.
• Employed Gamma correction for preprocessing to enhance damage area visibility.
• Calculated grayscale and HOG features for unit cells, combining them into grayscale-weighted HOG (GHOG) feature patterns.
• Input GHOG features into a specifically structured CNN.

Main Results:

• The GHOG-based CNN method demonstrated significantly improved performance compared to traditional HOG methods.
• The GHOG-feature-based CNN technique showed flexibility and effectiveness, achieving similar accuracy to deep learning methods using raw data.
• The method proved to be more efficient, reducing machine time and hardware requirements.

Conclusions:

• The proposed GHOG-feature-based CNN method offers an efficient and effective solution for highway pavement damage detection.
• This approach reduces computational load, making advanced detection accessible on common hardware.
• The method's foundation on grayscale, a feature with physical meaning, suggests potential for future detailed damage analysis.