A Performance Improvement Strategy for Concrete Damage Detection Using Stacking Ensemble Learning of Multiple...

Area of Science:

• Civil Engineering
• Computer Vision
• Machine Learning

Background:

• Semantic segmentation networks offer pixel-level concrete damage detection.
• The performance of individual semantic segmentation networks is often constrained.
• Improving concrete damage detection requires advanced methodologies.

Purpose of the Study:

• To propose a stacking ensemble learning method for concrete crack detection.
• To enhance the performance of semantic segmentation networks in identifying concrete damage.
• To evaluate the effectiveness of ensemble learning in improving concrete damage detection accuracy.

Main Methods:

• A database of 500 concrete crack and spalling images was created and split into training and testing sets.
• Five semantic segmentation networks (FCN-8s, SegNet, U-Net, PSPNet, DeepLabv3+) were trained and validated using a five-fold cross-validation approach.
• Softmax regression-based ensemble learning models were trained on predictions from the semantic segmentation networks.

Main Results:

• The best ensemble learning model showed performance improvements of 0.21% PA, 0.54% MPA, 3.66% MIoU, and 0.12% FWIoU compared to the best single network.
• Ensemble learning effectively aggregated predictions from multiple semantic segmentation models.
• The proposed method demonstrated superior concrete damage detection capabilities.

Conclusions:

• Stacking ensemble learning significantly improves concrete damage detection performance.
• Ensemble learning provides a robust strategy for overcoming limitations of single semantic segmentation networks.
• The developed method offers a promising advancement for automated concrete inspection and maintenance.