Research On Segmentation Algorithm Of Pavement Crack In Complex Environment

With the rapid development of China's economy,the mileage of China's highway has been growing gradually.By the end of 2018,the total mileage of China's highways has exceeded 4.85 million kilometers,which puts forward higher requirements for highway maintenance.Cracks are the initial symptoms of pavement damage.In order to ensure traffic safety,pavement cracks should be found and remedied in time.Automatic pavement crack detection device can shorten the time of manual maintenance and make the work safer.Because of these advantages,the device is preferable for the industry.Most of the automatic crack detection algorithms developed in recent years work well under the condition of clear crack image and simple background.However,these algorithms are difficult to meet the actual engineering requirements.The environment of crack is much more complicated in real life,due to many effects such as light and shadow.Therefore,the paper mainly studies the crack image segmentation algorithm,and providing relevant technical support for the subsequent development of more intelligent automatic pavement crack sealing machines.The paper is mainly carried out research work in the following aspects.1)This paper builds a real environment crack image dataset,which is called Crack-Data.It solves the problem that there are few public datasets for pavement crack image,and the object in the existing datasets is clear with simple background.The dataset is captured on four roads in two cities and contains a variety of complex backgrounds,such as shadows and zebra crossings.The corresponding crack image labels are made according to the format of PASCAL-VOC2012 dataset,in order to detect cracks by using the deep learning segmentation method.2)A pavement crack segmentation algorithm based on crack-FCN network is proposed.On the basis of FCN,the network structure has been improved,which solves the problem that the segmentation result is rough,the crack details are not obvious and there are discontinuities.The control experiment between FCN and crack-FCN on the crack-Data dataset shows that the segmentation result of the crack-FCN algorithm combines more crack information,enriches the crack details,and improves the performance compared with FCN.However,for the segmentation result of fine cracks under the illumination shadow,the discontinuity problem cannot be completely eliminated.In addition,FCN and crack-FCN networks have longer training time,slower recognition,and less efficiency.3)A method of pavement crack segmentation based on VGG-U-net network is proposed.It solves the problem that there is a gap,the training recognition time is long,and the efficiency is low in the crack segmentation result in a more complex background.It constructs two parts of the encoder and the decoder,which improves the accuracy of edge information prediction.By migrating and fine-tuning the trained VGG16 network model parameters,a new hybrid model VGG-U-net is constructed in combination with the classical U-net network.Compared with FCN,crack-FCN and U-net on the crack-Data,the results of experiment show that VGG-U-net not only recognizes the high segmentation accuracy,but also takes less training time,which is more conducive to the real-time process.