Research On Pavement Disease Identification Method Based On The Deep Learning

With the rapid development and improvement of China’s transportation system,road maintenance problems have become increasingly prominent.Pavement cracks,as the main manifestation of pavement diseases,realize rapid and efficient intelligent identification,which helps promote the development of national intelligent pavement quality management and lays the foundation for the construction of intelligent transportation.At present,road surface disease detection mainly relies on traditional manual detection.When faced with the detection of massive road surface diseases,it exhibits shortcomings such as strong subjectivity,missed and wrong detections,and low detection efficiency.Due to the diversity of pavement cracks and the complexity of the pavement environment,it also makes traditional image processing techniques difficult to recognize and low in recognition accuracy.In view of the above problems,the thesis takes pavement disease images as the research object,and studies the deep learning model of feature reduction and the convolutional neural network based on transfer learning to achieve automatic and accurate recognition of pavement cracks.The main research work of this article is as follows:(1)In view of the complexity of the road environment,multi-scale,diversity and the rich information contained in the pavement disease image,it is difficult to characterize the complete pavement crack characteristics.A variety of feature extraction methods were studied including texture description,external image description,and principal components Analysis description and wavelet singular value decomposition description.Based on the obtained feature quantities,a feature set representing pavement cracks is established.(2)In view of the redundant or irrelevant feature attributes existing in the high-dimensional features of pavement cracks,the Relief F and NCA algorithms are used to eliminate the irrelevant or redundant feature attributes in the high-dimensional feature set,and retain the feature quantities that are more sensitive to the prediction results.It analyzes the use of manifold learning t SNE to map the original features to the low-dimensional data space,find the low-dimensional manifold structure of the high-dimensional space,and realize the feature reduction.(3)Aiming at the problem that most machine learning models cannot self-learn effective feature information and the accuracy is not high,deep learning model recognition methods are studied.The establishment of DSAE,DBN and ML-ELM deep learning models and network parameter settings were discussed.According to the established model evaluation index,the recognition performance of the depth model based on feature reduction was compared and analyzed,and the performance evaluation of the road surface disease recognition model was completed.According to the experimental results,the combination of NCA+ML-ELM model is more accurate and more efficient in identifying cracks on the pavement,and its accuracy rate reaches99.95%.(4)In view of the small size of the target data set,the difficulty of complete characterization of pavement crack feature information and the difficulty of separate convolutional neural network training,a convolutional neural network based on transfer learning is implemented to realize automatic recognition of pavement cracks.The VGG-19,Inception-V4 and Res Net 152 pre-training models trained under big data are discussed,and some of the parameters and all parameters are selectively transferred to the target dataset.The convolutional neural network trained under various optimization algorithms is used to realize the automatic recognition of pavement cracks,which solves the problem of long training time and difficult adjustment of small-scale target data sets in pure convolutional neural network,and improves the pre-trained convolutional neural network.Accuracy on the target data set.According to the experimental results,in the training strategy 1,using the optimizer Adagrad or Rmsprop,the recognition accuracy rate of VGG-19 and Res Net 152 can reach 100%.