Research On Road Crack Recognition Algorithms Based On Deep Learning Multi-model Fusion

Using traditional image processing methods to recognize pavement cracks requires a lot of image preprocessing,which leads to inefficient,due to the multi-texture,multi-objective,weak signal and background complexity of pavement images,the recognition accuracy is not high.In order to solve these problems,this paper proposes a pavement crack recognition algorithm based on deep learning and multi-model fusion on the basis of domestic and foreign research.The algorithm fuses multiple neural networks to achieve accurate classification of pavement crack images.The main research work of this paper is as follows:(1)Analyse the difficulty of pavement crack identification,and set up a data set to include a large number of crack images with complex background in the data set,so as to improve the practicability of the algorithm.In order to increase the amount of data and improve the robustness of the model,real-time data enhancement method is used to transform the crack image on the CPU while training the network on GPU.(2)The classical neural network models LeNet-5,AlexNet and VGGNet16 are used to test on the pavement crack data sets,and the improvement strategies are proposed respectively.By analyzing the experimental results,it is concluded that the complexity of the model is lower than that of the crack image training set.It is necessary to further increase the network level in order to achieve better recognition effect.(3)The over-fitting phenomena occurring in the experiments using residual networks ResNet34 and ResNet50 are analyzed,and a new residual block improvement method is proposed.Dropout is added to the residual blocks and batch normalization and ReLU activation function are put before the convolution layer.The experimental results show that the method effectively reduces the over-fitting of the network.After that,the network depth is reduced and the Inception V3 model with increased width is used.The weight of the model is only about 1/150 of ResNet50.The training speed is fast and the generalization ability is strong,and the recognition rate is good without over-fitting.(4)A multi-model fusion algorithm for pavement crack recognition is proposed by comparing and analyzing a large number of previous experiments.On the basis of ResNet34,ResNet50 and Inception V3,the network structure and algorithm implementation flow of thefusion model are designed,and the recognition effect of the fusion model is better than that of the single model in the experiment.It is proved that the variance of the model can be reduced by processing data through several different models and then combining them before entering the full connection layer.Different models have different complexity and can better extract the generalization characteristics of the image when they are combined.