Research On Extraction Technology Of Urban Buildings Based On High-resolution Remote Sensing Image

As an important part of the high-resolution remote sensing image,the artificial surface features on the earth's surface,as an important bridge,connect the remote sensing satellite and human daily life.At the same time,the interpretation of buildings also plays a key role in population estimation and urban planning,early warning and assessment of natural disasters,and monitoring and utilization of land.However,due to the complexity of the scene of high-resolution remote sensing image,the shape of the buildings is different and affected by the light,shadow and other factors.If the traditional building detection method is used to extract the high-dimensional characteristics of the target buildings,it is very time-consuming and unable to mine.Faster R-CNN uses the shared convolution network to obtain the deep features of the original image,and then combines the regional recommendation network to generate the preliminary recognition results.Therefore,this paper takes Faster R-CNN as one of the methods of building extraction.However,Faster R-CNN itself can not achieve the effect of edge contour extraction when it does target detection,so this paper introduces the Level-Set method based on active contour model to improve the segmentation accuracy of buildings.In this paper,a high-resolution remote sensing image building extraction method based on Faster R-CNN and Level-Set is proposed.Faster R-CN-N algorithm can determine the location of the target building accurately,but it can not extract the outline of the building accurately.Traditional Level-Set algorithm is widely used in medical image processing,especially in medical image segmentation.Because of its high segmentation accuracy,the final curve can converge to the edge contour of the target,the author decided to apply this algorithm to the building extraction in this paper.Because the initial contour is interfered by human factors,the segmentation effect is not ideal when the target is occluded and the gray level of the target and the background are similar.In order to solve these problems,the author integrates the deep learning technology into the traditional Level-Set algorithm,and proposes a building extraction algorithm combining Faster R-CNN and Level-Set.Then,an improved Level-Set algorithm is proposed.When the Level-Set algorithm is used to extract the buildings in the remote sensing image,due to the objective reasons such as the illumination and the position of the buildings in the nature,the curve often stays on the shadow of the buildings rather than the outline of the buildings at the end of the evolution,which is under segmented,and the accuracy of the algorithm is not high enough;at the same time,the traditional Level-Set algorithm is strongly dependent on the setting of the initial outline The cut result will vary with the initial contour,and the Level-Set function needs to be reinitialized continuously in the process of iteration,which requires a lot of computation.In view of the problems in building extraction of the above-mentioned traditional Level-Set algorithm,the author adds a priori shape energy term to the traditional energy function,and at the same time accelerates the convergence speed by adding a penalty term.Finally,the author proposes an automatic convergence method based on "average distance",which makes the whole algorithm more intelligent.The experimental results show that the "Faster R-CNN+improved Level-Set" method proposed in this paper can accurately identify the buildings in the remote sensing image and has strong robustness.It not only solves the problem that the convolution neural network method can not extract the external contour of the target,but also promotes the application of active contour model algorithm in remote sensing image recognition.