Research On Road Extraction From High Resolution Images With Stroke Width Transformation And Convolutional Neural Network

As the main and basic way of transportation,the roads play a key role in the development of transportation system.Research on the extraction of road information is also of great significance in traffic management,including automatic road navigation,urban planning,and road monitoring,driverless vehicles,map updates,etc.,which are very important in industry and daily life.Road extraction is one of the most important tasks in modern transportation system,which is usually difficult,because the complex background of a remote sensing image,such as rural roads,has a heterogeneous appearance with large similar and low inter-class changes;urban roads,covered vehicles,pedestrians and the shadows of surrounding trees or buildings also increase the difficulty of this task.Although road information extraction from high-resolution images is a challenging research direction,but it also has great research significance.Based on the research results of other scholars,following researchs have been done in this paper,and the extraction of road information has been achieved successfully and effectively by this paper's method :(1)Remote sensing technology and the characteristics of roads in high-resolution images are summarized,and the theoretical knowledge of stroke width transform algorithm,convolutional neural network and mathematical morphology are introduced,which lays the theoretical foundation for the study of road information extraction methods in this paper.(2)Considering the strong consistency of road width,a method of extracting road information using a stroke width transformation algorithm is proposed.However,other features around the road in a image have a large impact on the extraction accuracy of the algorithm,which makes it very difficult to obtain high-precision road extraction results.In view of this,an improved stroke width transformation algorithm is proposed,which combines the advantages of mean shift and stroke width transformation.Firstly,in order to reduce the error extraction phenomenon of stroke width transformation algorithm,the preprocessed images are segmented by mean shift;Secondly,the pixels in the segmented image are divided into two categories by stroke width transformation algorithm,namely road and non-road,which can achieve the extraction of road information and obtain road images;Finally,the road images are optimized by mathematical morphology operation to make the obtained road information more accurate.The experiments show that the improved algorithm can accurately achieve the extraction of road information from high-resolution images,and the accuracy is better than using the mean shift and stroke width transform algorithms alone.(3)By studying the relevant theories of convolutional neural networks and analyzing its advantages in road extraction applications,it can be concluded that the structure of convolutional neural networks has the feature of hierarchicalization.And with its great learning and expression ability,road features can be extracted automatically in massive experimental data,which is also very suitable for road extraction in complex scenes and can solve the problem of poor extraction accuracy of stroke width transformation algorithm under complex road background.Therefore,the convolutional neural networks are introduced into the road extraction in this paper,and a road extraction method combining U-Net network and stroke width transformation algorithm is proposed.The first is to conduct the U-Net network training for the data set after data enhancement,and the road information is initially extracted by the trained network.Then,with the help of the road width and edge information,the stroke width transformation is used to process the initial extracted road image,and the shallow features of the road are used to eliminate the non-road part of the road image to a large extent.Finally,morphological processing is performed to optimize the road extraction results and the accuracy evaluation is also carried out.This method can realize the extraction of road information,make the extracted road edges smooth and avoid the appearance of interference and burr.