Cracked concrete surface classification in low-resolution images using a convolutional neural network

Evaluating the quality of concrete structures relies significantly on detecting cracks, as they impact the safety, usability, and longevity of the structure. Cracks appearing on the concrete surface serve as early indicators of structural damage, making their identification crucial for maintenance. In a manual inspection, the initial step involves outlining and documenting the crack's characteristics. However, the accuracy of quantitative analysis in manual assessments heavily depends on the expertise and experience of the specialist. As an alternative, we propose the use of automated image-based crack detection methods. Four Computational Machine Learning (CML) methods and one deep learning method for crack detection in concrete structures, have been examined k-nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree (DT), Artificial Neural Network (ANN), and Convolutional Neural Network (CNN). Each method was assessed for its ability to classify crack and non-crack images across different resolutions and training setups, aiming to enhance structural health monitoring through effective crack detection. This study aims to identify cracks in buildings using low-resolution images, and we trained and assessed KNN, SVM, DT, ANN, and CNN algorithms using various image sizes, including 50 x 50, 35 x 35, 25 x 25, 10 x 10, and 5 x 5 pixels. Upon analysis of the sample images at resolutions of 50 x 50 and 5 x 5 pixels, it was observed that the CNN classification approach consistently achieved the highest accuracy levels, ranging from approximately 99-95%. In comparison, the other four techniques, namely KNN, SVM, DT, and ANN, yielded accuracy rates ranging from 89 to 91%, 98-90%, 97-89%, and 94-92%, respectively. These findings suggest that KNN, SVM, DT, ANN, and CNN algorithms maintain high accuracy even when applied to low-resolution images of 5 x 5 pixels, comparable to their performance with higher-resolution 50 x 50 pixel images.