| At present,the detection of asphalt pavement disease mainly relies on the engineering vehicle to drive along the road line,and to collect the pavement image by CCD camera,then the centralized analysis and processing is carried out.In the case of large data volume,it is very labor-intensive,especially in the face of the increasing expansion of asphalt roads,pavement disease detection technology need to be improved.Related research has attempted to solve the problem of pavement disease image classification by using traditional machine learning methods,but this type of method is mainly based on shallow models.On the one hand,the model has a simple structure and a small number of layers.The feature extraction mainly relies on manual completion.It is difficult to dig deep features,the expression ability of the model is limited,and the classification effect is not ideal.On the other hand,road types are diverse and complex,and the collected road images are easily polluted by lane lines,lights,shadows,etc.Conventional image preprocessing methods are difficult to adapt to such complex environments.With the rapid development of artificial intelligence,the intelligence of related technologies in the field of transportation based on deep learning has received a lot of attention.Considering that convolutional neural networks are especially good at processing image classification problems,this paper focuses on the CNN method for the classification and identification of asphalt pavement disease images.The main research contents are as follows:Firstly,an removal method for lane lines of the asphalt pavement image based on Mask-RCNN and improved Criminisi was proposed.To solve the problem that a large number of road images with lane lines easily affect the final classification effect of CNN.First of all,the algorithm use Mask-RCNN to detect the lane line region,remove the lane line by the detected mask,and then fill the blank area with the improved Criminisi algorithm,complete the lane line removal.The comparison experiment showed that the accuracy of the new data set processed by the proposed algorithm is improved about 3% on the AlexNet model.Secondly,an illumination equalization algorithm based on Mask uniformization and gamma correction was proposed.To solve the problem of poor quality of the collected pavement image,such as interference noise or uneven illumination.First of all,the initial illumination correction of the image is completed by Mask homogenization method,and then the gamma correction method is used to further adjust the brightnessof the image,and enhance the contrast.The illumination equalization is completed.The gamma coefficient is found in the custom gamma list by the dichotomy.Experiments showed that the new data after illumination equalization has better convergence on the AlexNet model,although the accuracy rate is not obvious.Thirdly,an image classification method based on CNN for asphalt pavement disease was proposed.In view of the limitations of traditional machine learning methods,CNN was introduced into the classification of asphalt pavement images.Focus on the four classic classification models of AlexNet,VGGNet,GoogleNet and ResNet,and improve or optimize the relevant network structure according to actual needs.Experiments showed that the performance of four models is: ResNet>GoogleNet>AlexNet>VGGNet,with an iteration of 10,000 times as an example.The classification accuracy on the test set is 99.75%,99.08%,98.67%,and 82.25%,respectively,classification results of CNN methods are significantly better than SVM methods.Conclusion: This paper effectively solves the problem of classification and identification of asphalt pavement disease images.Among them,the lane line removal method can better solve the lane line interference problem in classification;the illumination equalization method can obviously improve the quality of the pavement image and solve the problem of uneven illumination;ResNet works best in the classification of disease images on asphalt pavement. |
PDF Full Text Request