Unsupervised Deep Learning for Road Crack Classification by Fusing Convolutional Neural Network and K_Means Clustering

Abstract

Pavement crack detection, classification, and characterization are key components of automatic pavement monitoring systems. Most of the currently acceptable methods used for these purposes are less satisfactory, which resulted in the fact that manual condition surveys are still adopted by many departments, considering the overall cost of pavement assessment. Moreover, current advanced deep learning-based road crack classification approaches are primarily supervised methods, which reduced the efficiency and hindered the intelligent aspects of the methods. Considering these problems, this research conducted edge-cutting work in proposing unsupervised deep learning methods for pavement crack classification. The proposed method fused convolution neural network architecture with the K_means clustering algorithm to achieve unsupervised feature learning. Considering the simplicity aspect of the model, AlexNet is selected as the base CNN model to build the fused architecture. The constructed crack classification model was trained and tested on road crack images collected by both automatic vehicles and smartphones, which showed satisfactory performance. Meanwhile, comparative experiments of transverse, longitudinal, and alligator crack classification are conducted between the proposed method and traditional unsupervised method, which proved the superiority of the proposed method (average accuracies for transverse, longitudinal, and alligator are 0.806, 0.792, and 0.913, respectively).