Research On Remote Sensing Image Semantic Segmentation Based On Encoder-Decoder Convolutional Neural Network

Remote sensing image segmentation is the basis of understanding remote sensing image,and plays an important role in disaster prevention and mitigation,environmental governance and urban planning.In the past,traditional methods related to image segmentation were often used to process remote sensing images.Traditional image segmentation methods are very low in both precision and efficiency.In recent years,with the wide application of deep learning in the field of computer vision,especially the remarkable advantage of convolutional neural network（CNN）in feature extraction,the speed and precision of image segmentation have been greatly improved.Meanwhile,the model based on deep learning can realize convenient end-to-end training and learning.Therefore,it is of great significance to study and design a deep learning network for remote sensing image segmentation.The work done in this thesis is as follows:1.This thesis briefly summarizes the research background and significance of remote sensing image segmentation,as well as the status quo of research at home and abroad,and classifies and summarizes the existing remote sensing image segmentation methods.Then,the theoretical knowledge of deep learning and image segmentation,as well as the classic image segmentation model based on convolutional neural network are elaborated in detail.2.The network structure of codec is a typical image segmentation model based on deep learning.In the coding part,convolution,pooling（subsampling）and other operations are included.Among them,pooling is essentially a kind of information sampling.By subsampling,the image can be compressed to extract the advanced semantic features of the image.But there is a sampled signal that may completely degenerate into a different form,a phenomenon we call aliasing.This will cause the network to be unable to reconstruct the fine-grained details of the image at the decoding end.In order to reduce aliasing effect caused by subsampling,a low-pass filter layer is inserted before subsampling to filter the foreground and background information of the image adaptively.In addition,in the codec structure,the jump connection connects the encoder to the decoder,which reduces the loss of fine-grained spatial information caused by subsampling to a certain extent.However,the integration of features at different levels is ignored,and there may be semantic gaps.Residual fusion module is introduced into jump connection,which can optimize network training and reduce the difference of feature semantics at different levels.3.Contextual information is crucial to image segmentation,especially in remote sensing images with large scale changes.The acquisition of contextual information is based on multi-scale features.In the middle part of the codec structure,a parallel module CEM composed of void convolution with different sampling rates is used to extract the multi-scale information of features and integrate the context information related to the target.In addition,due to the large difference of background and the imbalance between foreground and background in remote sensing images,the multi-scale information has some interference of redundant information.We use force mechanism to suppress the interference of redundant information and highlight the usefulness of the target.The effectiveness of the proposed model was verified by using attention-guided context Extraction and testing on Deep Globe Road Extraction and Inria Aerial Image Labeling datasets.