Employing histogram of oriented gradient to enhance concrete crack detection performance with classification algorithm and Bayesian optimization

• Published in: Engineering failure analysis Vol. 150; p. 107351
• Main Authors: Sun, Zhen, Caetano, Elsa, Pereira, Sérgio, Moutinho, Carlos
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2023
• Subjects: Concrete bridges; Crack detection; Histogram of oriented gradients; Machine learning; Vision-based methods; Histogram of oriented gradients; Crack detection; Concrete bridges; Vision-based methods; Machine learning
• ISSN: 1350-6307; 1873-1961
• DOI: 10.1016/j.engfailanal.2023.107351

•A crack detection method for concrete bridges is presented using a binary classification algorithm and the histogram of oriented gradient (HOG) feature.•HOG feature can capture the physical shape information of cracks in the images, thus overcoming the black-box nature of deep learning methods.•Bayesian optimization is used to tune hyper-parameters to improve classification accuracy and efficiency by employing the Gaussian process.•SVM classification algorithm is used to achieve an accuracy of 94.38%, outperforming CNN, Adaboost, KNN and Naïve Bayes algorithms. This paper develops a crack detection method for concrete bridges using a binary classification algorithm, which utilizes the histogram of oriented gradient (HOG) feature to effectively and efficiently capture the crack characteristic. The method is implemented in three steps: collecting images with and without cracks, HOG feature calculation as input variables, and crack detection with a binary classification algorithm. Moreover, the influences of different HOG parameters are investigated to better capture the cracks' characteristics. The classification algorithm is adopted as a support vector machine (SVM) model, and Bayesian optimization is employed to select hyper-parameters in the model training. The method is verified with available datasets of images with and without cracks from concrete bridge decks. The corresponding performance is compared to that of other methods, such as convolutional neural network (CNN), Adaboost, K-Nearest Neighbor (KNN), and Naïve Bayes, evidencing higher accuracy and computation efficiency.