Research On Road Crack Identification Based On Full Convolution Neural Network

In recent years,China’s highway infrastructure construction is becoming more and more perfect.The traditional road maintenance method is carried out by inspection personnel using human eye observation and manual recording,which has the disadvantages of low detection efficiency and strong subjectivity,and is no longer suitable for the current new situation of road maintenance.The crack is the most common form of pavement diseases,in view of the existing crack detection algorithm in large-scale applications especially wide region and many roads algorithm under complicated environment stability of the crack detection based on the analysis of the existing literature,on the basis of in the field of semantic segmentation,a crack of automated segmentation algorithm is put forward,mainly working content is as follows:1.Production of road crack data set.In view of the lack of sample data of crack labels,a large scale original image set of pavement cracks is collected,Labelme tool for pixel level manual marking and then extended in the form of data to enhance the image number,After removing all negative samples,9825 labeled images were obtained.Finally,pavement crack data set was constructed for subsequent model training.2.In the early processing stage of pavement crack image data set,In view of the obvious features of crack edge,a new gray-scale algorithm of crack image based on adaptive linear weighting is proposed during the gray-scale processing of original crack image.The algorithm takes the edge of the image as the main feature information.It not only reduces the amount of model training data but also retains the edge information of cracks effectively.3.Complete the segmentation task of road cracks based on FCN and U-Net models.Firstly,adjust two kinds of network input size for crack image segmentation tasks,In the construction of U-Net model,in order to cope with the characteristics of small segmentation target and large segmentation background,the traditional convolution operation method is improved,The improved residual convolution is used to enhance the feature extraction ability of shallow network,and a semantic segmentation network suitable for crack segmentation is obtained.Then the two kinds of networks are trained on the self-made data set and the corresponding weight files are saved.Then two common semantic segmentation evaluation indexes and verification sets are used to evaluate the preservation model.Compared with FCN,the segmentation effect obtained by improved U-Net training is better.4.A novel road crack segmentation method based on convolution-deconvolution feature fusion for integral nested network SP-UNet is proposed.Based on Encorder and Decorder architecture,backbone network composed of residual convolution as encoder enables the model to extract richer feature information,In order to solve the problem of feature redundancy in the coding stage,the SE module of channel attention mechanism is integrated in the jump connection to assign specific values to different channels and increase the weight of segmentation targets.and to solve the problem of feature redundancy in the coding stage,The SE module of channel attention mechanism is integrated on the jump connection to assign specific values to different channels and increase the weight of segmentation targets.Secondly,according to the morphological characteristics of fractures,the location attention mechanism is introduced into the model to increase the ability of the model to capture the long-distance dependence relationship,improve the model’s "receptive field",and achieve better results in fracture segmentation with linear topology.The SP-UNet model constructed in this paper is also trained and evaluated using road fracture data sets.Secondly,according to the morphological characteristics of cracks,the location attention mechanism is introduced into the model to increase the ability of the model to capture long distance dependence.The model’s "receptive field" is improved to achieve better results in fracture segmentation with linear topology.The SP-UNet model constructed in this paper is also trained and evaluated using road fracture data sets.5.Finally,accuracy rate and recall rate,were used as quantitative evaluation indexes to verify the contribution of each module in improving the performance of crack segmentation.Experimental results show that compared with channel attention mechanism,the location attention mechanism module has obvious advantages in improving the performance of crack segmentation,and has good application and promotion and this point has good application and promotion.