Deep learning-based concrete defects classification and detection using semantic segmentation.

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Deep learning (DL) offers potential for accurate, objective infrastructure damage detection in structural health monitoring (SHM).
• Key challenges include limited defect image datasets and selecting appropriate DL network architectures for real-time applications.

Purpose of the Study:

• To develop and evaluate DL models for concrete crack and spalling detection using a novel dataset.
• To address limitations in existing defect image databases and DL network depth selection for SHM.

Main Methods:

• A diverse dataset of 4087 concrete crack and 1100 spalling images was created.
• Convolutional Neural Network (CNN) classifiers (VGG19, ResNet50, InceptionV3) were used for defect identification.
• Encoder-decoder models (U-Net, PSPNet) with various backbones (VGG19, ResNet50, InceptionV3, EfficientNetB3) were developed for semantic segmentation.

Main Results:

• InceptionV3 achieved 91.98% accuracy for defect classification with the RMSprop optimizer.
• EfficientNetB3-based U-Net yielded the best crack segmentation (95.66% F1-score).
• InceptionV3-based U-Net excelled in spalling segmentation (89.43% F1-score).

Conclusions:

• The developed DL models and dataset significantly advance automated visual damage detection in SHM.
• Specific CNN architectures demonstrate high efficacy for both classification and segmentation of concrete defects.
• This research provides a foundation for more accurate and accessible real-time structural health monitoring systems.