Research And Implementation Of Pavement Crack Detection Algorithm Based On Multi-network

With the rapid development of road transportation system,driving has become the mainstream mode of travel.The normal operation and safety of the road transportation system is the basis for the normal operation of society.Road maintenance can effectively prevent the further development of pavement diseases,improve the service life of roads,and enhance the comfort and safety of drivers.Crack,the most common pavement disease,is the most direct and effective indicator to evaluate the pavement condition.The maintenance department can generate pavement assessment reports on the basis of the pavement crack detection results,which can guide the subsequent road maintenance work and provide effective support for the safety of road traffic.In the modern era of rapid development of information technology,deep learning has been studied quite deeply in the field of pavement crack detection.However,due to the diversity of pavement backgrounds,crack types and construction interference,it is difficult for a single model to completely fit the complex and variable pavement characteristics.The identification of pavement cracks is still a difficult problem to solve.Therefore,this paper proposes a multi-network pavement crack detection algorithm.The main content of this article is as follows:1.The purpose is to achieve a pavement crack detection system based on deep learning,using pavement data collected by CiCS cars.It is supposed that the trained model can be directly applied to road maintenance companies in the actual production.It is expected that this detection system can smoothly replace the existing pattern identification,and provide information for road traffic safety,and therefore realize effective social benefits.2.The primary crack model is obtained by training the crack dataset with deep convolutional neural network,which can play an important role in crack detection with the Dice coefficient of 80.8%far better than the existing pattern identification.The white crack model is obtained by training through white crack dataset,which is much better than the crack primary model in the detection of white cracks.The lane line model is obtained by training through the mark and line dataset,which has a unique advantage in the integrity identification of the road marking lines.Each single model has its own practical significance.3.A new multi-network model is proposed,which includes the addition and subtraction of the main model and the auxiliary model,and the introduction of the decision network.The purpose is to explore the combination of the main model and the auxiliary model to improve the overall identification effect of crack detection.Firstly,by combining "crack primary model and white crack model",the identification results are added in matrix dimension.The index of white crack dataset is improved without affecting the index of the basic crack detection dataset.Secondly,"crack primary model-lane line model" is used to improve the identification of cracks by eliminating the error identification of mark line cracks.Finally,by introducing the decision network,the multi-layer perceptron is used as the decision network to intelligently combine the recognition results of multiple models,and the Dice coefficient reaches 82.38%,which effectively improves the detection effect of pavement cracks.The experimental data show that model summation,model subtraction and the introduction of decision networks in the multi-network model are beneficial for the detection of pavement cracks.