COMPUTER VISION-BASED SURFACE CRACK DETECTION IN CONCRETE PAVEMENTS: KEY METHODS, ARCHITECTURES, AND CHALLENGES

Abstract

Concrete pavement crack detection is a critical component of infrastructure maintenance and structural health monitoring because surface cracks directly influence pavement durability, safety, and service life. Traditional manual inspection methods are labor-intensive, subjective, time-consuming, and inefficient for large-scale transportation networks. Consequently, computer vision-based crack detection techniques have emerged as a promising alternative for automated pavement assessment. This review critically examines recent advances in computer vision-based surface crack detection in concrete pavements, focusing on preprocessing methods, segmentation strategies, convolutional neural network architectures, transformer-based models, and multimodal sensing approaches. The study highlights the transition from conventional image processing and handcrafted-feature machine learning techniques toward deep learning-based semantic segmentation frameworks capable of achieving high pixel-level accuracy. Widely used architectures such as U-Net, DeepLabv3+, SegNet, ResNet, EfficientNet, MobileNet, and transformer hybrids including SwinUNet and CrackFormer are comparatively analyzed in terms of accuracy, computational efficiency, robustness, and deployment suitability. The review further discusses benchmark datasets, evaluation metrics, real-time implementation challenges, dataset imbalance, environmental variability, and the limitations of current systems in practical pavement management applications. Recent developments in attention mechanisms, RGB–infrared fusion, stereo vision, and lightweight embedded models are also explored. Finally, the paper identifies future research directions emphasizing multimodal datasets, efficient transformer architectures, domain adaptation, and integration of crack detection with automated structural condition assessment systems.