Research On Road Fracture Recognition Based On Deep Learning

The traditional method of road fracture recognition mainly uses image processing technology to extract the feature of crack image.This kind of method has limited ability to extract the deep level feature of the crack image,and the recognition effect is difficult to achieve the expected effect.With the development of computer vision and machine learning,it is possible to simulate the expression of the hierarchical structure of the human brain.Many deep neural network models are proposed by the designer.The model adopts multi-level and nonlinear network structure to extract image features from layers.The extracted features are highly abstract and discriminative,and are suitable for complex multi classification problems.Since advanced learning has been put forward,it has attracted many domestic and foreign scholars to pay great attention to it and become a trend in the field of pattern recognition.In this paper,the deep learning method is applied to crack identification,and the following three tasks are carried out.First,the application status of road fracture identification technology at home and abroad,and its progress in the field of image processing and deep learning are introduced.The concept,core idea,working principle and characteristics of deep learning theory are expounded,and the common deep learning model and working principle are introduced in detail.Second: Aiming at the low recognition problem caused by complex background interference such as low contrast,obstacle and random noise,a method of highway crack detection and recognition based on deep convolution neural network is proposed.In the coiling layer,the traditional convolution process is replaced by the expansion convolution process.In the polymerization layer,the threshold aggregation method is used instead of the maximum aggregation and the mean aggregation method,and the collection features are classified accurately with the decision tree classifier.The experiment shows that the improved CNN network can improve the convergence performance by 18.36%,and the error recognition rate can be reduced by 26.74%,as compared with the traditional ones.Therefore,the improved expansion convolution neural network has better generalization ability and robustness.Third: in view of the poor recognition rate caused by illumination factors such as road illumination,debris reflection and scattering,a multi-channel convolution neural network based on SDHR is proposed to detect and identify highway cracks.First,the SDHR operator is used to extract the global and local texture features of the cracks.Then the convolution neural network is input to convolution and pooling operation.Then the local and global features of the output are concatenated,and feature fusion and dimension reduction are realized through PCA algorithm.Finally,softmax classifier is applied to achieve accurate classification.Through 10 model cross experiments,the correct recognition rate of this method is 97%,and the average correct rate is 85%.Therefore,the method has good effect on the expression of fracture characteristics,and the accuracy of fracture recognition is greatly improved.