Comparative Analysis of Machine Learning and Deep Learning Classifiers for Crack Classification

In historical structures, cracks on the exterior are vital indicators of possible structural damage. Historical structures face an elevated susceptibility to structural deterioration due to the influence of natural disasters and indirect human factors. Efficient and cost-effective methods for detecting cracks are of utmost importance in the monitoring of the structural health of the historical buildings, as they facilitate the prompt identification and resolution of potential problems. This paper conducts a comparative analysis for concrete crack detection, employing both deep learning and machine learning approaches. Four machine learning classifiers, namely SVM, decision tree, random forest, KNN, and extreme learning machine (ELM), are compared against pre-trained deep learning models such as VGG16, VGG19, and InceptionV3. The comprehensive evaluation is performed on a well-known historic building crack dataset, which is balanced through the application of data preprocessing techniques. The models are validated on the processed dataset and their performance is compared. Among machine learning classifiers, random forest emerges as the top performer, while in the realm of deep learning classifiers, VGG16 takes the lead. Specifically, VGG16 achieves a remarkable accuracy of 92.14%, surpassing the performance of random forest, which achieves an accuracy of 66.92%. These results highlight the effectiveness of VGG16 in handling the intricacies of historic crack detection, showcasing its superior performance compared to random forest in the context of accuracy.