Research On Road Recognition Intelligently From High Resolution Remote Sensing Image

In recent years,high resolution remote sensing technology has become an important means of accurate access to surface information with the rapid development of multi-platform remote sensing technology and digital image processing technology.As an important part of basic geographic information,road network has play a very important role in vehicle navigation,land planning and so on.Traditional road extraction techniques based on high-resolution remote sensing images are usually accompanied by artificial visual interpretation,Though it can accurately obtain road network information,the efficiency and the degree of intelligence is low,also it can not update road network automatically.Therefore,it is of great meaning to carry out a high-performance and intelligent road network extraction method.At present,the road extraction methods,which includes two major steps:road binary contour map extraction and road center line refinement.Due to the complex context feature and spectral character of high-resolution remote sensing images,which leads to automated extraction of road networks have some shortcomings.(1)Based on the pixel level of the road extraction methods,prone to "salt and pepper"and "empty" phenomenon.While based on the object level of the road extraction algorithm,prone to "sticky" phenomenon,the existence of these two phenomena to a large extent affected the road network of the follow-up optimization steps.(2)In the initial extracted road image,the existence of non-road area which affects the uniqueness and completeness of the road target,and the state of art non-road area removal algorithm has the problem of poor robustness and universality,and needs to use more features and make complex parameter settings.(3)The existing road center line extraction result still have some shortcomings,one of the reason is that the perspective of the image itself,due to high-resolution image of the road there are vehicles,building shadows,etc.,prone to "fracture".Second,from the perspective of the centerline extraction algorithm,the existing road center line extraction algorithm is easy to produce "burr",the fitting is not accurate and so on.In view of the above research objectives,this paper mainly completed the following three aspects.(1)In this paper,a multi-level feature fusion algorithm considering pixel-level spatial features and object-level spectral features is developed from the perspective of feature extraction at the bottom of the image,which effectively compensates for the lack of single-level information and improves the accuracy of road extraction from the perspective of feature extraction,and improve the existence of the "salt and pepper" and the "sticky"phenomenon.(2)In order to solve the existing non-road area removal algorithm,there are too many feature combinations and parameter settings.In this paper,a new shape description operator is designed,and automatic filtering of non-road area is realized based on OSTU automatic threshold method.(3)Aiming at the phenomenon of "salt and pepper" and fitting inaccuracy of the existing road centerline algorithm,the tensor voting algorithm in computer vision is introduced to solve the above problems effectively,and the automatic connection of the fault road is realized,which improve the accuracy of road extraction result.In general,this paper proposed a robust multi-feature fusion method and remove the non-road area automatically,and improves the integrity level of road network,finally forms a complete set of high-resolution remote sensing image road extraction method.In this paper,we designed a number of experiments,and compared with the state of art road extraction algorithm,No matter from the qualitative point of view or quantitative point of view,the algorithm has achieved better results,demonstrated the effectiveness and robustness of our algorithm.