Design and Performance Evaluation of Infrastructure Crack Detection Using UDLV3+
Lee Jong-Hyun
Department of Computer Engineering
Honam University
Directed by prof. Lee Sang-Hyun
This study aims to design and evaluate the performance of deep learni...
Design and Performance Evaluation of Infrastructure Crack Detection Using UDLV3+
Lee Jong-Hyun
Department of Computer Engineering
Honam University
Directed by prof. Lee Sang-Hyun
This study aims to design and evaluate the performance of deep learning-based segmentation models for more precise and efficient crack detection in concrete-based infrastructure structures. In infrastructure maintenance, cracks are among the most critical defects, and accurately detecting early-stage cracks plays a key role in determining the overall safety and lifespan of the structure. However, conventional visual inspection methods and crack gauge-based measurements have clear limitations in terms of objectivity and quantifiability, and are prone to subjective inconsistencies depending on the inspector’s expertise. To address these issues, this research proposes an automated crack detection method using artificial intelligence technology, specifically image-based deep learning segmentation models.
To achieve this, four models were selected for comparative evaluation: the widely adopted U-Net, DeepLabV3+, Swin Transformer, and the proposed hybrid model combining U-Net and DeepLabV3+ (UDLV3+). The models were compared based on their architectural features, training stability, inference speed, and quantitative performance indicators. The proposed UDLV3+ model was designed to combine the lightweight structure and fast convergence of U-Net with the multi-scale feature extraction and boundary refinement capabilities of DeepLabV3+, aiming to maximize crack detection accuracy while maintaining computational efficiency suitable for real-time applications.
Experiments were conducted using a dataset of over 40,000 concrete surface images classified into Positive (with cracks) and Negative (without cracks) categories. Preprocessing and augmentation techniques such as mask generation, normalization, resizing, and horizontal flipping were applied to improve generalization performance during training. The models were trained for 50 epochs, and performance was evaluated using metrics including IoU, Dice coefficient, inference speed (FPS), computational complexity (GFLOPs), and the number of parameters (Million Parameters).
As a result, the proposed UDLV3+ model achieved the highest detection performance with an IoU of 0.523 and a Dice coefficient of 0.554. Furthermore, it maintained a computational cost of only 1.43 GFLOPs and 26.18M parameters while achieving an inference speed of 519.70 FPS, demonstrating a successful balance between accuracy and efficiency. The U-Net model, with a Dice coefficient of 0.472 and an inference speed of 1978.91 FPS, showed extremely fast processing capabilities and was found to be suitable for lightweight or edge device-based systems. In contrast, DeepLabV3+ effectively captured cracks of varying shapes and scales, but its high computational cost (10.26 GFLOPs) and large parameter count (39.63M) make it less suitable for resource-constrained environments. Despite being a state-of-the-art Vision Transformer-based architecture, the Swin Transformer showed poor accuracy with an IoU of 0.125 and Dice coefficient of 0.216, indicating difficulty in distinguishing cracks from background noise and blurry boundaries.
Additionally, analysis of training loss curves showed that both the U-Net and UDLV3+ models exhibited rapid loss reduction in early epochs followed by a stable convergence phase. The hybrid model, in particular, demonstrated both fast convergence and low final loss values, confirming its robustness in training and strong generalization potential.
In conclusion, the proposed UDLV3+ model outperformed individual models in terms of accuracy, computational efficiency, and inference speed, demonstrating a level of performance suitable for practical application in real-world crack diagnostic systems. Furthermore, this study presents a viable AI-based solution for various application environments such as automated infrastructure maintenance, unmanned inspection systems, and Edge-AI-based real-time diagnostics. These findings suggest a foundational contribution to the advancement of smart city infrastructure management and AI-integrated structural health monitoring systems.