Evaluating of KNN, Random Forest, and ResNet18 for Identifying Cracks in Historical Buildings

A key element of structural damage monitoring and restoration efforts in old structures is crack detection. Conventional manual inspection techniques lack scalability, are labor intensive, and are susceptible to human error. This work investigated the performance of K-Nearest Neighbors (KNN), Random Forest, and ResNet18 models in autonomously finding cracks from the Historical-Crack18-19 dataset. The classes were balanced by increasing the dataset, which consists of, 3139 non-cracked images and 757 cracked images before training. Classification reports, confusion matrices, and predictions based on samples helped assess the models. The tests produced findings demonstrating that ResNet18 far outperformed Random Forest and KNN. Comparatively, KNN is 82% and Random Forest is 88%; the accuracy of ResNet18 is 99%. ResNet18, which is based on deep learning, had the best accuracy, Recall, and F1-Score metrics. It also had a way to tell the difference between surfaces that are cracked and those that are not. These findings show image convolutional neural networks (CNNs) are better at finding cracks, which means they could be used to keep an eye on old buildings in the real world. We will use sophisticated deep learning architectures and domain adaptation methods in future work to increase the model's resilience on many datasets even further.