Research On Road Information Extraction Technology Based On Remote Sensing Images

With the continuous development of remote sensing in my country,it has become more and more convenient to receive remote sensing images on the ground.The information carried in the obtained remote sensing images has created convenient conditions for the development of all walks of life in society.Remote sensing images play an important role in urban and rural development planning,environmental monitoring,and smart transportation.As the most common ground objects in remote sensing images,road information is of great significance in real life.After the painstaking research of scholars at home and abroad,many algorithms for traditional image road extraction have been proposed one after another,and many results have been achieved.However,the results of traditional road extraction algorithms in remote sensing image road extraction are difficult to satisfy,and there are problems such as low accuracy and low degree of automation,which cannot meet the needs of road extraction from remote sensing images with complex and changeable ground objects.Therefore,the road extraction technology of remote sensing image urgently needs to develop in the direction of faster,more accurate and more intelligent.At the same time,as one of the safety signs of road information,zebra crossing plays a vital role in traffic safety,and also attracts the attention of many scholars.The complex background,redundant information,and diverse types of remote sensing images make road extraction and zebra crossing detection also face the problem of low detection accuracy.In this thesis,the deep learning method is used to design and build a road extraction model and a zebra crossing detection model,and based on this the remote sensing image road information analysis system is designed and implemented.The main contents and merits are as follows:(1)Proposed an improved U＿Net road extraction method.First,the encoder uses VGG16 network structure to replace the original U＿Net encoder structure.Then,feature compression activation module(SENet)is added after each encoder and decoder block to enhance network feature learning ability.Finally,the loss function combined with Dice loss function and dicclassification cross entropy loss function was used to reduce the sample imbalance problem in road extraction task.The road detection data set of Saertu District of Daqing city is constructed.The results on The Massachusetts Road dataset and the Saertu District dataset show that the improved algorithm can effectively improve the Road extraction results.The accuracy,recall rate,F1-Score and m Io U evaluation indexes of the proposed method in the Massachusetts Road test set reached 82.5%,77.8%,80.0% and 82.1%,respectively.The accuracy,recall rate,F1-Score and m Io U evaluation indexes of the proposed method in the Saertu District test set reached 81.3%,75.0%,79.4% and 81.6%,respectively.In the test image,it has better recognition effect on the intricate road.At the same time,in order to adapt the network to the prediction of large remote sensing images,the cropping and splicing method that ignores the edges is used to avoid the problem of obvious image splicing traces.(2)Zebra crossing detection method of four-channel network model is proposed.This method is improved on the basis of the network model proposed in research content(1).Zebra crossing data sets were captured from Ole Earth and labeled with Labelme.First of all,the dataset is coverted to grayscale images.Then,the data set is denoised by mean filtering and median filtering.Finally,the dynamic threshold is used to binarize the de-noised image.The dynamic threshold is the average of the difference between the grayscale image and the denoised image.The results obtained after morphological processing are used as the fourth channel of the image,and the four channels of image red,green,blue and binarized are fused into the improved U＿Net network for network training.In zebra crossing test set,the accuracy,recall rate,F1-Score and m Io U evaluation indexes reached 79.8%,75.0%,77.5% and 80.1%,respectively,and the experimental results were improved compared with the original network experimental results.(3)To check the validity of the research content of this thesis and the practicability of deep learning in remote sensing image processing,this thesis designs and implements a remote sensing image road information analysis system,which is suitable for road supervision departments to "build-management-maintenance-transport".The system fully uses deep learning algorithms and combines py Qt5 and My SQL technologies.Through demand analysis,modules such as login,registration,road detection and area calculation,zebra crossing detection and damage analysis are designed.Finally,by simulating the actual operation of the user,the system is tested by the black-box testing method,which further proves the practicality and effectiveness of the algorithm proposed in this thesis.