Research On Semantic Segmentation Of Road Cracks Based On Deeplabv3+

Maintaining and rehabilitating roads not only extends the life of the pavement,but also increases the comfort and safety of vehicles during travel,while the costs of road operation are reduced.The detection and repair of pavement cracks is a very important task in the process of pavement maintenance.The timely repair of cracks can prevent large-scale pavement reconstruction and reduce the likelihood of road accidents.With the development of deep learning,the use of neural networks will have a faster speed and higher accuracy than traditional manual when processing pavement crack images.Therefore,this thesis uses a convolutional neural network model to semantically segment and categorise road cracks to facilitate efficient crack repair work,reduce labour costs and improve the quality of work of road maintenance departments.The main research elements and innovations accomplished in this thesis are as follows:(1)Expansion and segmentation of the dataset.The existing road crack data set was expanded and subdivided into single cracks into single horizontal and single vertical,cracked turtles into cracked block and cracked dense network,and cracked repairs into cracked horizontal and cracked vertical.A Deeplabv3+ model incorporating the dense module and the attention module was trained and experimentally tested on the expanded dataset as the basis for the network.By comparing and analysing the crack segmentation result maps,the shortcomings of the model that still exist in the field of crack segmentation are summarised as a later study.(2)An improved road crack segmentation method based on the Deeplabv3+ decoder.In response to the problem that the Deeplabv3+ model with fused dense module and attention module has incomplete crack edge segmentation and local area fine cracks cannot be extracted effectively in the crack segmentation test results,this thesis designs a feature fusion module and a feature map slice module in the decoder.The feature fusion module is based on the idea of multi-scale fusion of U-Net networks,and fuses the shallow features of different resolutions extracted from the backbone network Xception with the deep feat ures after up-sampling,and then introduces the feature map slice module to perform slice amplification and feature extraction operations on the fused features,so as to improve the segmentation ability of the model for crack edges and small cracks in loca l areas.The experimental results show that the improved network model achieves an MIo U value of 85.2%on the road crack test set,which is 1.1% higher than the original model,and the segmented crack edges are more complete and some of the fine cracks can be extracted effectively.(3)An improved road crack segmentation method based on Deeplabv3+ lightweighting.In view of the current problems of large number of model parameters and computational volume,and the complex network structure,this thesis lightens the process.The lightweight model ECANet is introduced and fused to the end of each Flow block of the backbone feature extraction network Xception to reduce the complexity and computational burden of the network by enabling cross-channel information interaction without affecting the model performance;in addition,the normal convolution in the dense module Dense ASPP is replaced with a depth-separable convolution,which enables the model to improve both the attention to the crack information while reducing the number of parameters and computational effort of the model.Ultimately,the lightened and improved model achieved an MIo U of 85.3% on the test set,a reduction of 25.2M in the parametric volume of the model,a reduction of 39.8G in the computational volume,and an increase in FPS from 4.5to 15.7.(4)Design and implementation of road crack segmentation software.This thesis designs a road crack semantic segmentation software,which is based on Py Qt5 tool and can run cross-platform,and it mainly contains two modules of crack segmentation and data aggregation.The improved network model of this thesis is firstly embedded into the software,and then the crack image to be segmented is input,and the model will perform segmentation and classification processing operations on the input image,and finally the segmented crack result map and the crack classification summary information are displayed in the visual interface of the software.The improved model proposed in this thesis has significantly improved the segmentation effect of crack edge features and small cracks in local areas,and after the lightweight processing,the model can achieve real-time segmentation and data summarisation of cracks in the designed visualisation software,thus improving the efficiency of road maintenance personnel and saving labour costs.