| With the continuous improvement of highway maintenance tasks in China,accurately classifying road cracks and recommending appropriate preventive measures has become the focus of the current transportation industry.At present,the method that relies on deep learning to classify road cracks requires labeling the image in advance,which will also consume a lot of human and financial resources.Moreover,the existing unsupervised image classification methods have defects such as poor general applicability and low classification accuracy.Therefore,a study on an unsupervised pavement crack classification method incorporating deep learning was launched in this paper.An unsupervised classification model combining convolutional neural network and cluster analysis was proposed in this paper,which introduces unsupervised algorithms into deep learning.And the model was applied to the field of pavement crack classification to make up for the shortcomings of the existing classification methods.First,in this paper,the most typical AlexNet model and K＿means clustering algorithm were used to design the model.The features of the pavement crack images were extracted through the AlexNet network,and then the extracted deep feature vectors were used as the input of the K＿means algorithm to cluster.The output of the clustering was used to replace the real label for backpropagation and update the weight parameters of the convolution.These two processes were performed iteratively in order to achieve the purpose of unsupervised classification of images.Then,the convolutional network model for feature extraction and the algorithm to guide the clustering process were improved respectively,and the improved two parts were fused to achieve the purpose of optimizing the unsupervised model.The AlexNet was optimized from two aspects of network structure and model training.The K＿means algorithm was replaced by the adaptive fast peak clustering improved in the paper to guide the clustering process.The improved unsupervised image classification model has improved the classification effect of crack images compared to before,and to a certain extent,it has realized the automatic classification of cracks in pavement images.Finally,in order to verify the universality of the unsupervised image classification model proposed in this paper to other image data sets,four datasets commonly used in the field of image classification were selected for classification experiments,and the results were compared with several algorithms that currently perform relatively well on unsupervised image classification tasks.The comparative analysis results prove that the unsupervised classification model proposed in this paper has better performance than the existing unsupervised methods on different data sets.A large number of experiments show that the unsupervised classification model proposed in this paper can realize the classification of crack pictures without label information.It also eliminates the process of deep learning crack classification method that requires tedious manual labeling.The improved unsupervised model improves the average accuracy of crack images by 4.7% compared with the model before the improvement,and the classification accuracy of transversal cracks,longitudinal cracks and alligator cracks have 83.6%,82.1% and 74.5% respectively.It also proves that the model strengthens the automatic classification of pavement cracks to a certain extent,and provides the possibility of reducing the financial cost of pavement maintenance. |
