Research On Building Extraction Method Of High Resolution Remote Sensing Images Based On Semantic Segmentation And Instance Segmentation

The research of building extraction of high resolution remote sensing image(highresolution image for short)is of great significance in social and economic construction,urban planning and cartography.Due to the anisotropy of buildings and the interference of background features,the extraction results of high-resolution images based on traditional algorithms are generally deficient in accuracy and lack of edges.Therefore,it is urgent to develop a high-precision automatic extraction model for high-resolution images of buildings.In recent years,with the great progress of hardware equipment and the arrival of the era of big data,the deep learning algorithm has been developed at an unprecedented speed.Deep learning algorithm has the advantages of high accuracy and strong robustness in image processing.High resolution image extraction method based on deep learning can achieve higher accuracy.This paper summarizes and analyzes the related research on building extraction of highresolution image at home and abroad,and expounds the building extraction method principle of deep learning convolutional neural network.In order to further improve the accuracy of the semantic segmentation of buildings,a multi-task learning semantic segmentation network based on contour features is proposed.In addition,to obtain the results of the mask and frame extraction of buildings in high-resolution images at the same time,an instance segmentation based high-resolution image building extraction method is proposed.The main work of this paper is summarized as follows:(1)The principle and structure of classical semantic segmentation models are analyzed,including FCN,u-net and ResUnet models.Considering the unique contour features of building objects,a multi-task learning network based on contour features is proposed.With the help of multi-task learning,this model can simultaneously use the local information and global information of the image to add the auxiliary task of contour feature extraction in the building segmentation task,so as to improve the extraction accuracy of building edge by using the contour feature to constrain and adjust the parameters of the segmentation model.(2)The difference between deep learning object detection method and semantic segmentation method is analyzed,the principle of RCNN series algorithm is introduced in detail,and the structure and principle of Faster RCNN model is emphatically analyzed.Furthermore,the Faster RCNN model was used to realize the object extraction of high resolution image buildings.(3)On the basis of semantic segmentation and object detection,a building extraction model based on instance segmentation is proposed to improve the practical value of building extraction model.The Mask RCNN model was used to extract the building object in the highresolution image,and the segmentation Mask and target bounding box information of the building object were output at the same time.In addition,the accuracy of the two output results was compared with the above semantic segmentation results and the target detection results.In this paper,five kinds of high-resolution image data sets of buildings are collected and sorted out,and the corresponding annotation and sample expansion are made on the data,which are used for the training and testing of building extraction model.The experimental comparison on the same test set shows that the proposed multi-task learning network segmentation model based on contour features is significantly better than the classical u-net series of semantic segmentation networks.The model of building instance segmentation based on Mask RCNN was significantly higher than the Faster RCNN model in target detection results,and it could simultaneously output semantic segmentation and target detection extraction results,which further improved the practical value of the model.