Research On Remote Sensing Semantic Segmentation Algorithm Based On Lightweight Neural Networks

Recently,with the continuous development of satellite technology,high-resolution remote sensing images have become the main data source for earth observation and analysis on a global and regional scale.Semantic segmentation of high-resolution remote sensing images aims to interpret the images by segmenting them into semantically meaningful objects and assigning each part with a predetermined tag,and has been widely used in various remote sensing tasks.Compared with other scenarios,remote sensing scenarios have higher requirements for the lightweight and real-time performance of the semantic segmentation algorithms.In the context of the limited processing resources of space-borne satellites and low-latency operation requirements,how to achieve the balance between the high segmentation precision and realtime efficiency of the semantic segmentation algorithms has important research value and guiding significance.Based on the ISPRS high-resolution remote sensing semantic segmentation datasets,this thesis explores the lightweight semantic segmentation algorithm models from the perspectives of super-pixel segmentation pooling and dual-range context aggregation.The main research contents are as follows:First,from the perspective of model lightweight,lightweight design is carried out from the encoder and decoder structure of the backbone network Link Net.Besides,by densely connecting dilated convolutions with different dilated rates,a lightweight dense dilated convolution pyramid is proposed,to obtain denser multi-scale features at a lower computational cost.Finally,a super-pixel segmentation pooling layer is introduced into the model to refine the segmentation results while achieving end-to-end semantic segmentation.The model achieved58.10% and 67.21% m Io U on the ISPRS Vaihingen and ISPRS Potsdam datasets respectively,while ensuring fewer parameters and lower computational costs.Second,from the perspective of lightweight self-attention mechanism and up-sampling,two linear units are applied to replace the key and value of the self-attention to reduce computational cost and complexity.After that,a lightweight dual-range context aggregation module is designed,which captures the local features and global context extracted by convolutions and self-attention mechanisms by introducing a dual-branch structure,and applies the re-weighting method to achieve feature aggregation with high efficiency.Finally,a highefficiency linear up-sampling module is designed based on the multi-layer perceptron,which significantly reduces the number of parameters and computational cost of the decoder structure.This model achieved 59.46% and 68.04% m Io U on the ISPRS Vaihingen and ISPRS Potsdam datasets respectively.Compared with other comparison methods,this model achieves a better balance between segmentation accuracy and computational efficiency.In summary,two lightweight semantic segmentation algorithms based on super-pixel segmentation pooling and dual-range context aggregation proposed in the thesis can both achieve high-efficiency and high-precision semantic segmentation in high-resolution remote sensing scenarios.