| Road extraction has been a research hotspot in the field of remote sensing and computer vision in recent years,and it is also a key step in many applications,such as intelligent transportation,urban planning,GIS update and so on.However,the proportion deviation between road and background in remote sensing image is large,and the existing methods are difficult to use the semantic context in the image,which limits the accuracy of road extraction results.In view of the above problems,this thesis studies the automatic road extraction and post-processing optimization of road extraction results,and puts forward a relatively complete set of automatic road extraction and optimization model.The main contents of the model are as follows:（1）A-LinkNet road automatic extraction networkIn view of the high resolution of remote sensing image,the massive loss of road information caused by the traditional deep convolution neural network in the coding process,and the instability of network training caused by the use of dice loss function,which reduces the performance of the model,this thesis proposes a new type of road automatic extraction network,which integrates the spatial pyramid pool structure and residual structure of road extraction network（Atrus linknet,A-LinkNet）.The network is designed in the following aspects:firstly,on the basis of LinkNet network,the spatial pyramid pool module（ASPP）is added between the encoding and decoding stages to obtain multi-scale features and alleviate the problem of road information loss caused by pooling operation in the network encoding stage;Secondly,a new loss function（67）（69）（80）is designed。The loss function optimizes the dice loss function and linearly combines it with the cross entropy loss function to make（67）（69）（80）can alleviate the large proportion difference of road background in high-resolution remote sensing images,maintain a stable gradient descent direction,improve the stability of the loss function and avoid falling into the local optimal solution.（2）Remote sensing image road optimization model combining A-LinkNet and short-range conditional random fieldAiming at the problem that fully connected conditional random fields（Full Crf）has the phenomenon of over smoothing the road in the field of road extraction and can not effectively optimize the road shape,this thesis proposes a secondary optimization model for road extraction from high-resolution remote sensing images,short range conditional random fields（SRCRF）.The specific work is as follows:firstly,the fully connected neighborhood structure is optimized to K-Neighborhood structure to alleviate the excessive smoothing phenomenon.Secondly,the K-Neighborhood mean field reasoning algorithm proposed in this thesis is used to solve the reasoning problem of SRCRF,so as to quickly infer the best label corresponding to the observation pixels,so as to complete the accurate optimization of the road.Aiming at the task of automatic road extraction,this thesis proposes a set of solutions that take into account road extraction and road optimization.In the scheme,a new road extraction network A-LinkNet and a secondary optimization model SRCRF for the road extraction results of remote sensing images are designed and proposed.The two models in this scheme have the characteristics of low coupling and high robustness.The combination of the two models greatly improves the accuracy and flatness of road extraction results from remote sensing images.In Zimbawe dataset,A-LinkNet+（67）（69）（80）compared with the best result of the comparison network,the F1value is increased by 4.42%.On this basis,through the secondary optimization of SRCRF,the F1 value of road extraction results is increased by 2.05%;In Cheng dataset,A-LinkNet+（67）（69）（80）compared with the best result of the comparison network,the F1 value is increased by 5.57%.On this basis,after the secondary optimization of SRCRF,the F1 value of road extraction results is increased by 2.68%. |
