Portal de Periódicos da CAPES

A comparison study of semantic segmentation networks for crack detection in construction materials

2024; Elsevier BV; Volume: 414; Linguagem: Inglês

10.1016/j.conbuildmat.2024.134950

1879-0526

Zhongqi Shi, Nan Jin, Dongbo Chen, Dihao Ai,

Asphalt Pavement Performance Evaluation

Automated crack segmentation from digital imagery is crucial for infrastructure health monitoring via non-destructive evaluation techniques. However, accurately capturing complex, multi-branch crack patterns remains challenging. This paper presents a comparative study of four state-of-the-art deep learning models – UNet, FPN, PSPNet, and DeepLabV3 – for crack segmentation from video data. Various encoder backbones, including ResNet, ResNeXt, SENet, and DPN are examined under manual and automatic compression loading conditions. Cracks were induced in cement-based materials via a uniaxial compression loading system equipped with synchronized imaging, containing diverse crack phenomena ranging from single, discrete cracks to complex crack networks with branches and intersections. Segmentation datasets were manually labeled for model training and testing. Experiments were conducted to evaluate crack segmentation performance using four widely-used metrics: Intersection over Union (IoU), F1-score, Recall, and Precision. Results showed that UNet paired with SENet as the encoder backbone achieved the best overall performance by IoU, F1-score, and Precision for manual loading, while FPN leveraging a DPN68 encoder achieved the most balanced and robust tracking of intricate crack propagations. This work benchmarks state-of-the-art solutions and analyzes failure cases, guiding model design and continued research priorities for persistent challenges in crack monitoring tasks.