Automatic Pavement Defect Detection and Classification Using RGB-Thermal Images Based on Hierarchical Residual...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Accurate recognition of complex pavement conditions is crucial for infrastructure maintenance.
• Existing deep learning models often require significant computational resources and training time.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) for efficient and accurate pavement condition recognition.
• To enhance model performance by integrating an attention mechanism with RGB-thermal imaging.

Main Methods:

• An improved residual structure-based CNN was designed for lightweight classification.
• An attention module was embedded to optimize spatial and channel information processing.
• RGB-thermal images were utilized as input data.
• Gradient-weighted class activation mapping (Grad-CAM) was employed for result visualization.

Main Results:

• The proposed model achieved a maximum prediction accuracy of 98.88% using RGB-thermal input with an attention mechanism.
• The attention mechanism improved the model's focus on image details and channel utilization.
• The model demonstrated fewer parameters, reduced training time, and higher recognition accuracy compared to state-of-the-art deep learning models.

Conclusions:

• The developed lightweight CNN with an attention module offers a superior solution for complex pavement condition recognition.
• The integration of RGB-thermal imaging and attention mechanisms significantly enhances classification performance.
• The model provides an efficient and accurate alternative to existing methods for pavement analysis.