High-resolution Image Building Extraction Based On Convolutional Neural Network And Model Migration

Building extraction was an urgent problem in many remote sensing tasks.In low and medium resolution images,many buildings existed in the form of residential areas.In high-resolution images,the separation between buildings was clear.And the color,texture,structure and other characteristics of the roof were more prominent.It provided convenience for building extraction.This article studied from the following three aspects: the scale,model applicability and regularization in building extraction based on convolutional neural network(CNN).CNN with the single-scale was difficult to meet multi-scale building extraction.The article presented a new method of building extraction based on VGG16 to this problem.Firstly,the original images were downsampled at different scales.Then the features of buildings at different scales could be extracted.Meanwhile,in order to reduce the number of network parameters,the fully convolutional upsampling was used to replace the fully connected layer in the original VGG16 model.Due to the imaging time,illumination,sensor and other factors,remote sensing images were diverse.It causes the applicability of the model between different source data.To solve this problem,this article used the FastPhotoStyle algorithm.In this way,the difference between the two datasets with the same resolution was reduced.On the other hand,with the combination of semi-supervised learning and active learning,the new sample annotation was obtained.It not only reduced manual annotation,but also improved the accuracy of building extraction in model migration.The boundary of the building extracted by the CNN was not enough to reflect the unique geometry of the buildings.So it was necessary to regularize the boundary of the building extraction results.The minimum number of vertices algorithm could be used to regularize the right angle,diagonal and arbitrary angle building.But the algorithm couldn't automatically determine the type of building.In this article,a method for judging the type of building was proposed.By automatically identifying the type of building,the corresponding regularization method was selected.