Research On Pavement Crack Detection Based On Convolution Neural Network

In recent years,with the rapid development of deep learning technology,various highperformance neural network models have been proposed and widely used in various fields.The application of convolutional neural network（CNN）to intelligent detection of pavement cracks is not only one of the current research hotspots,but also of practical engineering significance.Data set is the key to train a CNN model,and it is also the main factor affecting the effect of crack detection.However,the traditional supervised learning method needs a large number of labeled data sets,which has some shortcomings.On the one hand,the labeling process consumes a lot of labor;on the other hand,because cracks have no fixed shape characteristics,their labels are prone to manual deviation.To solve these problems,this thesis carries out the research on pavement crack detection based on convolutional neural network models,and puts forward an improved semi supervised learning method combined with pseudo labeling and self-correction.This method takes the data set as the research object,aims to obtain high-quality data set and high-performance neural network model,and carries out the research in the following aspects through semi supervised learning method: reducing the labeling cost of crack detection data,defining the classification label of crack image data,determining the labeling method of target objects in crack region location,and separating crack pixels and their context information in crack semantic segmentation.The main work of this thesis is as follows:（1）Through pseudo label learning,each single class of pavement cracks and backgrounds is divided into multiple ordered subclasses,which expands the generalization ability of a neural network model and improves its accuracy of crack detection.Experiments on pavement crack images verify the effectiveness of the proposed method.（2）Aiming at the inconsistency of crack data labels,a self-correction technology is proposed,which corrects the errors in the original labels and improves the quality of data labels.（3）Integrate multiple models to detect pavement cracks in stages,that is,first locate the crack area in the road image through the image classification model Dense Net combined with the sliding window or directly use the target detection model Faster R-CNN,and then use Deep Labv3+ to segment the crack area semantically,identify the crack,obtain the crack pixels and other information,and calculate the crack length and other parameters.The results show that the proposed method can reduce the false detection rate and reduce the workload of manually labeling the pavement crack data set.