Optimizing CNN for pavement distress detection via edge-enhanced multi-scale feature fusion

Area of Science:

• Computer Vision
• Machine Learning
• Civil Engineering

Background:

• Automated road crack detection uses deep learning but struggles with complex backgrounds and noise.
• Existing methods lack robustness and generalization, impacting classification accuracy.

Purpose of the Study:

• To develop a novel deep learning approach for robust road crack detection.
• To enhance the accuracy and stability of road damage classification models in challenging environments.

Main Methods:

• Integration of traditional edge detection techniques (Sobel, Prewitt, Laplacian) with deep convolutional neural networks (DCNNs).
• Proposed Edge-Enhanced Multi-Scale Feature Fusion (EE-MSFF) mechanism for multi-scale edge information extraction and fusion.
• Utilized multi-scale receptive fields to capture local and global crack features, improving focus on damaged regions.

Main Results:

• Achieved 88.68% classification accuracy on the complex-background RDD2020 dataset.
• Attained 99.5% accuracy on the Concrete_Data_Week3 dataset with minimal background interference.
• Ablation studies confirmed performance improvements from integrating multi-scale edge features.

Conclusions:

• The EE-MSFF mechanism effectively mitigates background noise and enhances crack region focus.
• The proposed method demonstrates improved robustness and generalization for road damage classification.
• Edge-enhanced deep learning offers a promising direction for accurate and stable road inspection.