Sparse Villages Based On Knowledge Distillation Research And Application Of Building Segmentation

With the continuous development of deep learning and remote sensing satellite imaging technology,the segmentation effect of remote sensing data is becoming more and more accurate.However,models with high-precision effects are often accompanied by huge resource effects,relying on huge computing power and more memory,and cannot be used in more scenarios.Therefore,lightweight models have been receiving more and more attention in recent years.This thesis studies the lightweight of remote sensing image segmentation model.Aiming at the problem that the remote sensing building segmentation needs to consume a lot of memory and computing resources,the thesis takes lightweight as the starting point and focuses on the design of lightweight structures to carry out relevant theoretical and technical research.First of all,it analyzes the domestic and foreign status quo of remote sensing building segmentation,and describes the characteristics of village buildings,summarizes the related methods of lightweight segmentation network,uses the classic lightweight segmentation network,and describes the two data sets selected in this thesis And the reason for the selection;secondly,a network SGBUnet network based on hourglass-shaped depth separable convolution,using various lightweight techniques such as channel clipping and residual connection,and proposed in the lightweight segmentation network,the shallow layer is more important According to this point of view,the network structure was adjusted again,which improved the accuracy of the model to a certain extent;again,a novel dual-association self-distillation framework was proposed,which can extract more deep features of the model.into shallow structures.The dual-association self-distillation framework considers both the correlation between the spatial positions in the feature map and the correlation between channels for the connection between the deep features and shallow features of the network,and extracts this connection information from the deep network to the shallow layer.Migrating these association information from the deep network structure to the shallow network structure can help the model gain more knowledge at the feature distribution level,thereby improving the segmentation accuracy of the student model.Finally,combining the two proposed methods and related development techniques,a post-earthquake rapid reporting system is designed and implemented.Based on the above theoretical research,through experiments on WHU dataset and ISPRS Potsdam dataset,comparative implementation and ablation experiments,the experimental results of the lightweight model SGBUnet show that the model designed in this thesis compresses the number of parameters to 0.07 M,with only a small loss of accuracy.The experimental results of dual-correlation self-distillation verify that this structure has a certain effect on improving the accuracy of the model,and the MIOU indicators of the two models under different data sets have been improved to a certain extent,demonstrating the feasibility of the method.