Research On Pavement Crack Detection Algorithm Based On Deep Active Learning

With the rapid development of China’s traffic infrastructure,highway construction is the focus of the construction,pavement crack detection technology for road maintenance,traffic safety,traffic efficiency has an important role.In recent years,road crack detection technology based on deep learning has made great progress.However,the road crack detection based on deep learning still has some shortcomings.Firstly,the training process of deep learning requires a large amount of annotated data,which is time-consuming and laborious to annotate a large number of road samples.Secondly,in the process of collecting pavement data,the amount of data with and without cracks is unbalanced.Most of the road surface images are crack-free.If these crack-free images are judged by manual labeling,it will waste time.If all the collected road images are marked,the accuracy of the deep learning model will be affected by the data set with unbalanced data categories.The main research contents of this paper are as follows:In order to solve the problem of deep learning,a large amount of annotated data and imbalance of pavement image categories are needed in training.An algorithm framework combining convolutional neural network and active learning is proposed for pavement crack detection.In this article,the Resnet50 network performs two tasks.First,as a feature extraction network,the pavement image features are extracted.Second,the posterior probability distribution used to predict the corresponding categories of unlabeled images belongs to active learning.Firstly,the unlabeled road surface images were input into the Resnet50 network to generate the posterior probability distribution of the corresponding class of crack estimation for the unlabeled road surface images.Then the uncertainty of the unlabeled pavement image is calculated by the posterior probability distribution of crack estimation.Then,according to the uncertainty value of crack estimation,whether the sample contains rich information(features)is judged,and the data with rich information is selected to be labeled by experts.Finally,the Resnet50 network was fine-tuned using the road images annotated by the experts.An active learning method based on Gini coefficient is proposed.To summarize the existing active sampling strategies based on uncertainty,most of them use information entropy as the measurement of uncertainty.In addition,some papers introduce KL divergence and cross entropy as the measurement of uncertainty.These three methods are used in information theory to describe information richness and uncertainty.Therefore,we try to apply the same kind of Gini coefficient to the active learning sampling strategy to calculate the uncertainty of unlabeled road images.Experimental results show that the proposed deep active learning method for road crack detection can actively select road images with rich information from unlabeled road images to train Resnet50 network,and effectively reduce road image labeling work by about 80%.In the case of unbalanced pavement data set without labeling,the deep active learning method is better than the deep learning method of random data selection in terms of accuracy rate,recall rate and F1 value.In addition,compared with the minimum confidence,the active sampling strategy based on Gini coefficient proposed in this paper,the two active sampling strategies based on information entropy improved by 1.93% and 2.26% in the reduction of data annotation.In the training process,the convergence rate of accuracy and cross entropy loss of the deep active learning method based on Gini coefficient improved faster.