Pavement Crack Recognition And Lifting Method Based On Mixed Effects Model Prediction Method

After several rounds of large-scale construction,the focus of highway construction in China is gradually changing from "construction oriented" to "maintenance oriented" mode.In the daily maintenance process,effective identification of various pavement diseases is very important,and cracks are one of the most common and easily occurring diseases.Traditional pavement crack detection mainly relies on manual inspection.Although this method is relatively accurate,it requires a lot of manpower and material resources,but manual methods may have significant safety hazards.Therefore,how to build a more automated identification method is a matter of great concern in highway maintenance practice.In recent years,the practice of using the latest recognition methods based on digital images and deep learning for highway maintenance has gradually entered people’s perspective,which can effectively avoid many shortcomings of the above-mentioned manual inspection methods.However,in the process of identifying and measuring highway cracks,the accuracy of machine identification and later measurement needs to be improved.The main reason is that in the process of crack identification,it is easy to be interfered by some stains,which affects the identification of cracks and the calculation of crack length and width.This will also affect the evaluation of road quality.In view of this,this thesis will establish a statistical prediction model between machine identification values and manual identification values to further improve the accuracy of highway pavement crack information at a small cost.Specifically,this thesis will adopt the latest developed automated road crack identification system,which uses a convolutional neural network architecture suitable for road crack identification and is named YOLO deep learning algorithm by scholars.Using this system,the input crack image can be identified in a relatively short time,and information such as the type,length,and width of the crack can be provided.In order to prove the wide applicability of the model improvement method compared to the algorithm improvement method,this thesis will use two sets of automatic identification systems using the Yolov 4 and Yolov 5 algorithms to obtain two sets of lengths and widths given by the machine,and then use the model method to improve the accuracy of crack measurement.Note that in the actual situation,there are differences in factors such as road diseases,lighting,weather,and so on,so there is a sub overall situation.In view of this,this thesis innovatively introduces the mixed effect model into the field of crack identification and improvement.In order to increase the amount of data,this thesis uses the "leave one method" to divide the training set and the test set.Train the model in a training set,and calculate the percentage relative error value in a test set to evaluate the performance of different methods.In the evaluation of prediction models,this thesis compares the prediction results of linear mixed effect models with machine direct recognition and linear model prediction methods.In establishing the mixed effect model,this thesis uses the K-means clustering method.Research shows that the model based lifting method is superior to the machine direct recognition method on both sets of data,and the prediction method based on the linear mixed effect model is superior to the prediction method based on the classical linear model to varying degrees.This also indicates that the improved model based method has a wide range of applicability.At the same time,due to the low cost of model based methods,model improvement methods are of great significance for improving the recognition effect of a large number of crack images.