Electronic Theses and Dissertation

Road surface deterioration, including longitudinal cracks, transverse cracks, alligator cracks, and potholes, poses significant challenges to road safety, transportation efficiency, and maintenance management. although deep learning has substantially advanced automated road damage detection, existing approaches continue to face limitations in detecting small defects, handling class imbalance, adapting to cross-domain datasets, and maintaining computational efficiency for real-time deployment. this dissertation proposes a deep learning-based framework for accurate and efficient road damage detection by enhancing the yolov8 architecture. the proposed framework integrates a bidirectional feature pyramid network (bifpn) to improve multi-scale feature fusion, a channel-wise and spatial-wise hybrid attention module (chbam) to enhance feature representation, a mobilenet backbone to reduce computational complexity, and class-balanced cross-entropy (cbce) loss to mitigate long-tailed class distributions. the framework is trained using the road damage dataset 2022 and subsequently fine-tuned on a locally collected aceh road damage dataset to improve cross-domain generalization. furthermore, the proposed model is benchmarked against transformer-based detectors, including rt-detr, rt-detrv3, uav- detr, and codetr-swin, under identical experimental settings. experimental results demonstrate that the proposed framework achieves competitive detection performance while maintaining significantly lower computational complexity than transformer-based approaches, making it suitable for real-time road inspection applications. the integration of bifpn, chbam, mobilenet, and cbce effectively improves the detection of small and low-contrast road damages while enhancing robustness to domain shift between public and local datasets. this research contributes an efficient and scalable deep learning framework that bridges the gap between high detection accuracy and practical deployment, providing a viable solution for intelligent road infrastructure monitoring and maintenance.

Road surface deterioration, including longitudinal cracks, transverse cracks, alligator cracks, and potholes, poses significant challenges to road safety, transportation efficiency, and maintenance management. Although deep learning has substantially advanced automated road damage detection, existing approaches continue to face limitations in detecting small defects, handling class imbalance, adapting to cross-domain datasets, and maintaining computational efficiency for real-time deployment. This dissertation proposes a deep learning-based framework for accurate and efficient road damage detection by enhancing the YOLOv8 architecture. The proposed framework integrates a Bidirectional Feature Pyramid Network (BiFPN) to improve multi-scale feature fusion, a Channel-wise and Spatial-wise Hybrid Attention Module (CHBAM) to enhance feature representation, a MobileNet backbone to reduce computational complexity, and Class-Balanced Cross-Entropy (CBCE) loss to mitigate long-tailed class distributions. The framework is trained using the Road Damage Dataset 2022 and subsequently fine-tuned on a locally collected Aceh road damage dataset to improve cross-domain generalization. Furthermore, the proposed model is benchmarked against transformer-based detectors, including RT-DETR, RT-DETRv3, UAV- DETR, and CoDETR-Swin, under identical experimental settings. Experimental results demonstrate that the proposed framework achieves competitive detection performance while maintaining significantly lower computational complexity than transformer-based approaches, making it suitable for real-time road inspection applications. The integration of BiFPN, CHBAM, MobileNet, and CBCE effectively improves the detection of small and low-contrast road damages while enhancing robustness to domain shift between public and local datasets. This research contributes an efficient and scalable deep learning framework that bridges the gap between high detection accuracy and practical deployment, providing a viable solution for intelligent road infrastructure monitoring and maintenance.