Research On Key Technology Of Remote Sensing Image Semantic Segmentation Based On Neural Network

Semantic segmentation of remote sensing images is the first step in its specific application,which is widely used in autonomous vehicles,environmental monitoring and disaster warning.With the advent of big data and the substantial improvement of computer computing power,artificial neural networks have developed rapidly.Computers can simulate the human brain to process massive data,which greatly improves the processing efficiency of human beings.Deep convolutional network is of great significance for automatic semantic annotation of remote sensing images.In this paper,the convolutional neural network,which has been widely studied in the field of image segmentation in recent years,is applied to the field of remote sensing images.Based on the two requirements of accuracy and efficiency in remote sensing image segmentation tasks,we propose corresponding solutions.The specific research contents are as follows:Starting from the accuracy requirements of remote sensing images for segmentation,a duplex restricted network with guided upsampling is proposed.The network can be adaptively selected by detachable enhancement structure to achieve a trade-off between classification and localization tasks.To optimize the detailed information obtained by encoding,a ConcentrateAware Guided Upsampling module is further introduced to replace the traditional upsampling operation for resolution restoration.Besides,a Content Capture Normalization Module is used to enhance the features extracted in the encoding stage.Our approach significantly outperforms previous results with fewer parameters on two very high resolution(VHR)data,83.76%(vs82.42%)on Potsdam dataset and 85.87%(vs 82.74%)on Jiage dataset.Starting from the fitting efficiency requirements of neural networks for remote sensing images,a method to balance remote sensing images categories and optimize network training is proposed.The network adds the category balance coefficient α calculated by the dataset label prior information to the cross-entropy loss function,which can improve the network segmentation accuracy of the tail samples in the dataset.This strategy improves the Intersection over Union(Io U)of the tail category by 21.94% and 31.16% on the Potsdam and Jiage datasets,respectively.In addition,based on the category balance coefficient,a sample loss weighted λthat weights each image in the network training process is further proposed.The weighting method is generated by the training of weighted loss sub-network and attention mechanism module.The addition of this parameter improves the fitting efficiency of the network training process and stabilizes the network loss in advance.