An Urban Road Recognition Method That Combines Low-level Features And High-level Semantic Knowledge

With the advancing of the science,the resolution of remote sensing image rise continuously,this makes the traditional road recognition method that aiming at the medium and low resolution image inapplicable,finding new ways to recognize roads from high-resolution images is imperative.To solve the problem mentioned above,this paper proposes a road recognition method that aiming at high-resolution image,by researching the low-level feature and combining the extracted low-level feature and the high-level semantic knowledge,the urban roads can be recognized and extracted under the construction of the high-level semantic knowledge.Aiming at the common phenomenon of heavy traffic in the city road,this paper attracts feature following the attention diversion order of the human eye,by morphological operators of a certain size,the image is divide into the detail map that contains the small scale features like edge,cars and lane line,and the high scale map that contains the spectrum features and the shape features.Then the original image is segmented into super-pixels,by super-pixels,the scale of the object that will be merged in the next step is limited.To enhance and stand out the linear part of the detail map,the Frangi filter is executed,and based on the result,the tensor voting algorithm is executed to further integrate the linear elements that shear the same direction,like edges and cars under the condition of heavy traffic,thus the direction feature map with coherent direction features can be done.Based on the enhancement results of the direction feature,the original image is segmented into rough super-pixels to limit the scale of the targets that the object-oriented process will face later.Then under the guidance of the detail direction feature map,the super-pixels are merged,and the heterogeneity degree between super-pixels is calculated adaptively,thus the object-oriented segmentation can be done under the guidance of the region direction.The advanced object-orient segmentation algorithm can take full use of both the high-scale feature and the low-scale feature,and can achieve a bottom-up remote sensing image segmentation,and this algorithm also reduces the impact of the single feature of spectrum,on the condition of heavy traffic or other object that has strong direction features,the algorithm proposed in this paper can get a better segmentation result.Based on the object-orient segmentation result,the geometry and spectrum feature of each region can be extracted,and the relationship mapping table is built,based on the table,the low-level features are mapped into the high-level feature like roads,tree lawn and lane line.Then,by taking advantage of the abstract features that the high-level semantic contains to express the objects,based on the relationship of these three kind of objects and the semantic knowledge,the further judgment is done,and roads extraction results can be thinned and completed.The experimental results show that,the precision rate of the method that this paper proposes is high,thus this method can recognize and extract the urban roads precisely.