An Intelligent Method for Crack Classification in Concrete Structures Based on Deep Neural Networks

Identifying and examining the types of cracks in concrete structures is one of the challenging engineering issues. Detection of crack bifurcation is very important because it detects high-intensity surfaces in concrete structures. In this paper, a new architecture based on convolutional neural networks is presented for crack classification in concrete structures. The proposed architecture detected and classified crack bifurcation in less time and with higher accuracy than other conventional and authentic deep learning architectures. In this paper, the cracks in 12000 images of concrete structures were investigated by the proposed algorithm, which resulted in 99.3% accuracy in categorizing as non-cracked images, images with simple cracks, and bifurcated crack images. Moreover, the analysis of the confusion matrix showed an accuracy of 99.3% and a recall of 99.5%, which confirmed the proper performance of the proposed algorithm. The sensitivity analysis of the proposed algorithm also showed the need for proportionality between the number of data, the number of neurons in the fully connected layer, execution time, and the expected percentage of accuracy according to the application of the problem.