| dc.description.abstract | Pavement infrastructure forms the critical foundation of supply chains in both developing and developed nations. Maintaining optimal pavement performance is essential for effective transportation networks and economic functionality. Over time, pavement conditions deteriorate due to multiple interconnected factors, including climatic conditions, temperature variations, traffic loading, and seismic activity. Effective pavement management systems depend on accurate distress identification and assessment. Among various forms of pavement deterioration, cracking represents the most common and structurally significant distress type. While federal and local agencies continuously invest in data collection and pavement condition monitoring, traditional manual assessment methods remain labor-intensive, time-consuming, and prone to inconsistencies. This challenge necessitates the development of automated approaches for efficient and reliable pavement condition evaluation.
Recent decades have witnessed the development of diverse pavement crack detection methodologies, ranging from traditional image processing techniques to shallow machine learning algorithms and advanced deep learning approaches. Deep learning-based methods have demonstrated superior accuracy and enhanced generalizability compared to conventional image processing and shallow machine learning techniques. Nevertheless, these advanced solutions typically require significant computational resources, limiting their practical deployment.
In this study, we develop a comprehensive end-to-end framework for automated pavement crack severity classification, combining deep learning algorithms with image processing techniques to enable crack recognition, detection, and severity assessment. The proposed system emphasizes resource optimization, particularly regarding computational processing and memory utilization, to ensure practical deployment in pavement management applications.
The first study develops a lightweight deep learning solution for pavement crack recognition using a convolutional neural network architecture. The proposed model is evaluated across three public benchmark datasets and one private dataset, demonstrating effective crack pattern recognition in pavement imagery. Explainability analysis reveals that the model focuses on relevant morphological features for crack identification rather than relying on spurious correlations.
The second study addresses the challenge of developing a computationally efficient deep learning solution for pavement crack segmentation. We propose LiteCrackNet, a U-Net-based convolutional neural network that incorporates atrous convolution and multi-scale supervision for enhanced feature extraction. Evaluation across three public benchmark datasets demonstrates that LiteCrackNet outperforms existing state-of-the-art lightweight models, achieving superior precision-recall balance with reduced computational requirements. Cross-dataset evaluation further validates the model's enhanced generalizability compared to competing approaches. Notably, LiteCrackNet achieves state-of-the-art performance with only 0.043 million parameters, demonstrating exceptional parameter efficiency.
The final study implements image processing techniques to measure crack width from segmented images and categorize cracks into three severity levels: low, medium, and high. The classification system defines low-severity cracks as those with widths less than 3 mm, medium-severity cracks as those ranging from 3 mm to less than 6 mm, and high-severity cracks as those with widths of 6 mm or greater.
The proposed end-to-end framework delivers a computationally efficient solution for pavement crack classification while maintaining state-of-the-art performance levels, as demonstrated through comprehensive benchmark evaluations. | en_US |
