Research On Remote Sensing Image Semantic Segmentation Method Based On Convolutional Neural Network

With the great improvement of the earth observation capability of my country's highresolution series of satellites,the data contained in remote sensing images is becoming more and more abundant,and how to obtain accurate remote sensing image segmentation results has become a popular research direction.Semantic segmentation assigns semantic information to each pixel in the remote sensing image,and can extract information such as the location and category of the target in the remote sensing image.This paper mainly aims at the inaccurate segmentation of remote sensing image semantic segmentation due to the large difference in target scale and complex boundary,and optimizes and improves the U-Net network,which has excellent performance in the field of image segmentation,so as to improve the segmentation accuracy of remote sensing images.The main research work of this paper is as follows:First,this paper proposes a remote sensing image semantic segmentation method MCU-Net based on residual fusion and multi-scale context information.MCU-Net improves the ability of U-Net to acquire deeper features by using residual fusion module to deepen the network structure.Then,for the feature information between different levels,a top-down and bottom-up path is constructed to fully utilize the spatial information and semantic information contained in the shallow and deep features in the network by fusing features from different levels.In addition,for the feature information between the same levels,an enhanced atrous spatial pyramid pooling module is introduced,so that the output features have a wider range of semantic information.Secondly,on the basis of MCU-Net,this paper proposes DAMCU-Net,a semantic segmentation method of remote sensing images based on attention mechanism and edge detection.DAMCU-Net extracts global context information through the attention mechanism optimization module,and uses dense skip connections to fuse features to promote the network to recover more spatial details during the upsampling process,and use the FRe LU activation function to improve the network's ability to segment complex targets.For the edge information lost in the feature extraction process,an edge detection branch is added,and the feature information of the main path is supplemented by feature fusion to realize the optimization of the edge information.Finally,this paper validates the proposed methods MCU-Net and DAMCU-Net on two remote sensing datasets.The experimental results show that the segmentation accuracy of MCU-Net and DAMCU-Net is better than the existing methods such as FCN,U-Net,Res U-Net,U-Net++ and other algorithms,which can improve the problem of inaccurate target segmentation.