Release the Crack’CNN : Concrete Crack Detection using Convolutional Neural Networks

This study presents the development of an intelligent concrete crack detection system using Convolutional Neural Networks (CNNs) implemented in Python within a Jupyter Notebook. The research employs the SDNET2018 dataset, which comprises real-world images of concrete structures with varying crack patterns. The dataset poses challenges of class imbalance due to the unequal distribution of cracked and non-cracked instances. To address this issue and enhance model performance, the study incorporates Synthetic Minority Over-sampling Technique (SMOTE) for dataset balancing and GridSearchCV for hyperparameter optimization. The evaluation of three models, namely the initial CNN Model, the CNN Model with SMOTE, and the Optimized CNN Model with SMOTE, is conducted through cross-fold and hold-out validation. The optimized CNN Model with SMOTE demonstrates superior performance compared to the other models, exhibiting improved accuracy, F1 score, and ROC_AUC values. The optimized model achieves an accuracy of 0.8912, an ROC_AUC of 0.9301 and an F1 score of 0.8823, showcasing its effectiveness in accurately identifying concrete cracks. The findings from this study emphasize the potential of machine learning algorithms, particularly CNNs, in automatic crack detection for critical infrastructure.