Semantic Segmentation Of Oblique Photogrammetry 3D Data Based On Graph Neural Network

Oblique photography is a high-technology developed in recent years,which can obtain 3D oblique photography data quickly and efficiently.The semantic segmentation technology of remote sensing image based on deep learning has made great development in recent years and is widely used in urban planning,smart city and GIS construction.Compared with 2D remote sensing images,3D oblique photography data contains richer detail information,which can help to improve the semantic segmentation accuracy,but the segmentation of large-scale and large-scale 3D oblique photography data has not been reported yet.The thesis addresses the semantic segmentation of 3D oblique photography data,which has important theoretical significance and practical application value,and the main work is:Firstly,A specific framework of 3D tilt photography data segmentation is proposed to extract two kinds of data,Digital Orthophoto Map(DOM)and Digital Surface Model(DSM),from 3D oblique photography data.The DOM is the familiar airborne high-resolution remote sensing image,and the DSM is the corresponding altitude.So for the segmentation of 3D oblique photography data,it is only necessary to carry out the semantic segmentation of the corresponding high-resolution remote sensing image,and finally map the segmentation result back to the 3D oblique photography data to realize the segmentation.Secondly,This thesis designs a semantic segmentation neural network model SGA-Net(Self-Constructing Graph Attention Neural Network)for semantic segmentation of high-resolution remote sensing images in urban scenes.Based on the self-constructing adjacency matrix and graph attention network to effectively extract local spatial features of remote sensing images with elevation,the multi-view feature map is used to ensure the consistency of spatial geometry,and the channel attention mechanism is designed to further improve the performance of semantic segmentation.Thirdly,A semantic segmentation network model SGC-Net(Self-Constructing Graph Convolution Neural Network)for high-resolution remote sensing images in rural scenes is designed,SGC-Net adopts a network of Res Net50(50-layer residual net)variants for feature extraction,uses graph convolutional neural networks for global space and local space feature extraction and aggregation,and employs a channel attention mechanism to model each channel and enhance or suppress the relevant channels when performing different segmentation tasks.The experiments show that SGC-Net can effectively solve the semantic segmentation problem of high-resolution remote sensing images in rural scenes.