<![CDATA[This study offers a comparative assessment of the Support Vector Machine with Radial Basis Function Kernel and Extreme Gradient Boosting for automated concrete crack detection based on Histogram of Oriented Gradients feature extraction. Data comprised 40,000 RGB concrete images from an open-source Mendeley dataset; half were cracked and half were non-cracked. They processed through a preprocessing pipeline that includes the Poisson noise reduction and bilateral filtering techniques. Two approaches, holdout validation over several training/testing configurations (50:50, 60:40, 70:30, and 80:20) and systematic 5-fold cross-validation, were adopted for evaluation of the Wilcoxon signed-rank test for statistical significance and inference time for computational efficiency assessment. The experimental results indicate that SVM achieved a better holdout accuracy of 98.94% with the 80:20 configuration, while XGBoost achieved a cross-validation mean accuracy of 98.83% ± 0.0015. However, no statistically significant performance difference was revealed between the models according to the Wilcoxon analysis. Results indicated SVM excels at minimising false positives on undamaged surfaces, whereas XGBoost is better for identifying cracks, meaning that the choice of models used should depend on the application requirements, where applications require either the minimisation of false alarms or maximum sensitivity for detection in the case of structural health monitoring.]]>
Copyrights © 2025
