Road Crack Identification Method And Segmentation Method Based On Channel Attention Mechanism And Dense Connection Mechanism

Road cracks are a kind of disease caused by the external environment on the road,which may lead to more serious safety accidents if they are not repaired.Road crack detection is a detection technology that judges whether cracks are generated in the road through certain methods and means.With the rapid development of artificial intelligence,various artificial intelligence algorithms represented by deep learning have begun to be applied in the field of road crack detection.This paper introduces a road crack recognition network Dense Net-SE and its variant Dense Net-NSE based on channel attention mechanism and dense connection mechanism,and tests and analyzes the method and existing technologies.In addition,this thesis also introduces a variant road crack segmentation technology for road crack recognition.In this experiment,the standard training method and the generative confrontation method are used to train the same model,and the two methods are tested and analyzed.Specifically,the main research work of this thesis is as follows:（1）Analyze the technical difficulties of road crack recognition and road crack segmentation,compare the similarities and differences between the two methods,obtain the applicable scope,advantages and disadvantages of the two methods,and find the corresponding convincing methods for the two methods.Force data set.The use of data augmentation technology increases the sample size of the data set to improve the robustness of the model.（2）Using Alex Net,VGG-16,VGG-19,and Res Net-18 to conduct contrast experiments in the road crack dataset SDNET2018,and analyze the structure and training process of different models,for the follow-up Dense Net-SE and Dense Net-NSE provided a reference.（3）In the road crack recognition technology,this paper conducts a theoretical analysis on the network structure of the densely connected network Dense Net,and proposes a method to use the channel attention module to improve the Dense Net model,and obtains two new models Dense Net-SE With Dense Net-NSE.Experimental results and field test validation results show that Dense Net-SE and Dense Net-NSE can improve the generalization performance of the model,with higher reliability and mobility,and Dense Net-SE performs better than Dense Net-NSE.（4）In the road crack segmentation technology,this paper uses the standard training method and the generative confrontation method to train a full convolution model,and analyzes the characteristics and advantages and disadvantages of the two training methods respectively.Experiments have proved that the standard training method can converge faster and reach a better level.At the same time,the crack size calculation is completed,and the error is controlled within a good range.