Research And Implementation Of Satellite Image Semantic Segmentation Technology Based On Deep Learning

The accurate outline of building foundation marked by satellite map is of great significance in the work of geographic information survey,regional building planning and other fields.As satellite images are taken at a certain angle,the position of the building in the image deviates from the position of the foundation contour.Traditional image semantic segmentation method is based on the pixel value of the image,it can obtain the building image,but cannot obtain the foundation contour.At the same time,due to the lack of satellite image data sources,it is not easy to obtain the data set.The angles of satellite images in the data set are limited,leading the semantic segmentation model cannot be adequately trained.In this thesis,the following methods are carried out to solve these above problems,respectively.One problem is the insufficiency of satellite image data,The other is that the traditional image segmentation method cannot obtain the ground contour by segmenting the building image.In order to solve the problem of multi-angle satellite image sparsity,this thesis builds a generated convolutional satellite image model by using a video frame interpolation method,Moreover,it generates satellite images with new angles by satellite images in adjacent angles to expand the number of data.In the network,the pixel synthesis of interpolated frames is regarded as the local convolution of two input frames,it uses Encoder-Decoder structure to capture the local motion relationship and the pixel synthesis coefficient between frames to generates the convolution kernel for interpolation.It utilizes the separable convolution to separate convolution kernel,which dramatically reduces the calculation cost with the same kernel size.The model does not need to rely on optical flow information,and can well solve the scene of occlusion,blurring and sudden change of brightness in the input frame.Aiming at the deviation between the building image and the building foundation contour,which made the semantic segmentation algorithm can't get the contour correctly.The thesis proposes the Affine Trans Net semantic segmentation network,which uses the projection deviation correction module based on affine transformation to establish the transformation relationship between the building image and the foundation contour,so that the building image can be transformed into the foundation contour.Besides the network uses the expanded convolution module to improve the receptive field of the convolution kernel making better use of the semantic information of adjacent regions.In order to improve the sensitivity of the model to the tiny error in the image and the segmentation effect of the model,this network adopts the counter generation method to train the network.