Advanced machine learning approach for automatic crack detection and classification in concrete surfaces

Nowadays, one of the most commonly used construction materials is concrete. As a building material, concrete can be deformed under various conditions and cracks can form on this material. Depending on their condition and position, these cracks can pose serious hazards. Therefore, the automatic detection and classification of these cracks becomes a very important issue. The detection process, which is usually performed by manual observation, is labor intensive. In this research, a new machine learning method is proposed for automatic detection of cracks in concrete surface. The proposed method utilizes DenseNet201 based deep feature extraction approach. In addition, the model includes ReliefF-based feature selection and SVM-based classification phases. SDNET2018, an open access dataset, is used to test the proposed model. Both holdout cross validation and 10-fold cross validation techniques were applied for validation on this dataset. As a result of the test procedures, 93% classification success was achieved for 10-fold CV. The results obtained with the test procedures prove the success of the proposed method in automatic crack classification.