CONCRETE DAMAGE IDENTIFICATION FOR STRUCTURAL HEALTH MONITORING USING COMPUTER VISION
Authors/Creators
- 1. Queen's University Belfast, UK
- 2. University of the West Indies, Trinidad
Description
Detecting the early degradation of concrete structures can assist agencies in forecasting and planning future resource demands. Structural defects network SDNET2018 classes were used to create a custom binary convolutional neural network CNN classifier with physical crack damage predictive accuracy of 99.2%. Second, concrete's physical, chemical, and reinforcement damage classes, including a negative class, are stratified to generate a separate dataset using concrete defect bridge image CODEBRIM cropped image database to train a multiclass image classifier model with predictive accuracy of 80%. Comparisons were made based on the computational efficiency of a CNN architecture-specific classifier model and a ResNet50-retrained transfer-learning model predicting at 87.45% accuracy. These findings enable the generalisation and improvement of the damage assessment accuracy of concrete structures.