Abstract: Traditional remote sensing image target extraction methods mostly use visual interpretation or processing of pixel information, which is difficult to apply to complex scenes of high-resolution remote sensing images. However, the existing convolutional neural network semantic segmentation model may cause the extraction of target adhesion due to difficulty in achieving high precision. Aiming at this problem, this paper improves the instance segmentation model Mask R-CNN and proposes an efficient and accurate high-resolution remote sensing image building extraction algorithm. Firstly, convolution operation is added to the original feature extraction part of the Mask R-CNN to reduce the aliasing effect caused by upsampling. Then, a branch is added to the original mask prediction structure to improve the effect of mask prediction. Finally, train the improved network on the building dataset, the results show that the proposed method can accurately predict the top of each building independently, without target adhesion, and the mAP value is improved compared with the Mask R-CNN, which can effectively realize the refined extraction of remote sensing image buildings.

Key words: convolutional neural network, instance segmentation, Mask R-CNN, building, feature extraction