Research On Complext Pavement Damage Identification Algorithm Based On Deep Learning

As a huge public facility,the safety of the transportation system is related to the economic pulse and safe operation of the entire society.In order to maintain the safety of highway traffic,timely and comprehensive investigation of hidden dangers on the road surface is a very important task.Due to the large volume of road transportation and heavy road load in our country,a large amount of damage occur on the road every day.Cracks are the most common road damage.It does not only reduce road driving safety,but also poses a threat to the physical layer safety of the automated driving system,and even significantly reduce the road’s useful life.If pavement cracks are not repaired in a timely manner,it may cause a major safety accident.The traditional crack detection work mainly depends on the on-site inspection by the technicians of the relevant departments,yet this work is not only very difficult,but the results are also not satisfactory.Therefore,it is of great practical significance and application value to realize the automatic crack detection system in accordance to the real complex pavement situation.The current automatic crack recognition algorithms are mainly based on traditional digital image processing methods.The experimental verification of these algorithms is also mainly performed on the road image data with clean and ideal background information.The actual road image taken on the highway will not have only one crack,but will contain a large number of complex background interference items,mainly uneven lighting and shadows,as well as garbage,road texture,water stains,ruts,etc.,making the accuracy of the traditional crack identification method is greatly reduced or even fails.Among a large number of road background interference items,uneven lighting backgrounds and shadows are one of the factors that are ubiquitous in images and will disturb the crack recognition system.Aiming at the above-mentioned crack detection problem under complex pavement,based on the research on image recognition technology,this paper proposes a road crack recognition algorithm based on the illumination homogenization algorithm as a preprocessing method combined with deep convolutional neural network.After uniform light processing is performed on images with non-uniform illumination,shadows and other noise effects,the neural network is used to train and complete crack detection.The main work of this article is as follows:1)We research and analyze image enhancement algorithms that deal with the problem of illumination uniformity to improve the quality of training data and thus improve the accuracy of the model in crack recognition tasks.This article studies a variety of current image enhancement algorithms to deal with the problem of uneven illumination,including histogram equalization,homomorphic filtering uniformity algorithm based on the illumination-reflection model,an algorithm combining Retinex theory and gamma correction,and Mask dodging principle.We design comparative experiments,visually read and quantify the results of single image processing of different uniform light algorithms,and analyze the advantages and disadvantages of the algorithms.2)We develop a deep convolutional neural network image recognition technology and complete the design of crack recognition network.In this paper,the existing convolutional neural network is explained and analyzed,and the shortcomings are that the existing perception structure of the crack recognition structure are small,resulting in a low recognition rate.In view of the above problems,this paper designs a convolution structure based Deep neural network for extracting crack information;designs experiments to verify that the structure realizes a convolutional neural network with a larger perceptual field and deeper depth,which effectively improves the accuracy of crack identification.3)We test and tune the designed neural network with uniform light preprocessing.Test-train using datasets processed by different uniform light algorithms to compare the effects of different uniform light algorithms on the recognition performance of the neural network.We choose a suitable uniform light algorithm to further optimize the parameters of the neural network to obtain higher cracks Recognition rate.Experiments show that under the dataset of this subject,the crack image processed by Mask dodging principle has a more suitable average light intensity,which is better than the traditional image enhancement algorithm to uniformize the lighting effect on the actual road scene.Based on the data processed by uniform light,it is also beneficial for the neural network’s crack recognition rate.The crack recognition rate effect of the deep neural network based on Mask preprocessing proposed in this paper is less affected by noise such as shadows,and the final Dice similarity reaches 0.7937.Some research results of this article have been applied to relevant cooperation units.In view of the initial use effect,the crack recognition network designed in this paper based on uniform light pretreatment has good noise resistance and crack recognition rate.