Multi-scale Highway Disaster Image Segmentation Using Improved DeepLabV3+ Algorithm

With the vigorous development of China’s highway construction and the increasingly developed transportation,the requirements for highway safety inspection and maintenance are becoming higher and higher.At present,most areas are still dominated by manual monitoring,with high risk coefficient and low work efficiency.In recent years,artificial intelligence has developed rapidly,and deep learning has made breakthroughs in the field of computer vision.Deep learning is particularly outstanding in analyzing learning data,extracting features and identifying features,which solves many problems in the field of image processing.In this paper,a multi-scale feature extraction and lightweight semantic segmentation algorithm is designed based on deep learning,which can automatically identify road hazards and segment their specific locations.This algorithm can effectively improve the efficiency of road safety monitoring.The research content of this paper mainly includes the following aspects:Firstly,make the data set of highway disaster according to the research topic.Using the web crawler technology to make the primary data set,select the crawling image category relatively concentrated fault,collapse,fire,landslide as the research category of this topic.Then,the images with obscure features that could not be labeled were deleted,the image noise was suppressed by Retinex algorithm and histogram equalization,and the data set was expanded by random combination of rotation,scaling and translation.Finally,Labelme annotation tool was used for manual annotation.By using the above methods,the highway disaster data set used in this project contains 3452 images.Secondly,the improvement of DeepLabV3+semantic segmentation network is proposed.Encoderdecoder structure is adopted in the improved network.MobileNetV2,which has fewer parameters and lower accuracy than Xception network,is used as the feature extraction network in Encoder part,and the ASPP module of enhanced feature extraction is redesigned based on DenseASPP.A simple Encoder network is used to obtain multi-scale segmentation results.Furthermore,the improved structure,loss function and training strategy of DeepLabV3+semantic segmentation network are introduced in detail,and the experimental tuning is carried out on self-made highway data set.Experiments show that DASPP_MNV2_DeepLabV3+network designed in this paper can effectively segment and identify highway disasters.Miou is increased by 4.67%,MPa is increased by 3.33%,the segmented image is more complete,and the segmentation effect is significantly improved.