Research On Semantic Segmentation Method Of Remote Sensing Images Based On Deep Learning

With the continuous development of military UAVs,the requirements for the dynamic perception of the battlefield environment by UAVs are getting higher and higher.A remote sensing image semantic segmentation method that can be carried on military UAVs is indispensable to analyzing the battlefield environment more accurately.Using remote sensing image semantic segmentation,different areas and different targets in the battlefield environment captured by military UAVs through airborne cameras can be accurately segmented,which is conducive to accurate analysis of the battlefield and precise target localization,thus realizing full control of battlefield information from the visual level and providing guidance for the next battlefield reconnaissance or strike implementation.In this paper,based on deep learning,the Deep Lab V3+ network is used as the main framework to modify the model.The research is carried out in two aspects reducing the computational network parameters and improving the network segmentation accuracy,which are mainly studied as follows.First,the existing networks are pre-trained in the public dataset by studying the current semantic segmentation algorithms and their principles.The dataset and semantic segmentation network that are more suitable for this study are selected by combining the training results.We also make a self-built dataset for this study by collecting image data regarding the format of the public dataset and using it for subsequent experiments.Second,the Deep Lab V3+ network is improved by replacing the backbone feature extraction network with a lightweight segmentation network model that can be applied to mobile devices,optimizing the network structure,reducing the number of parameters,and improving the accuracy of the network while reducing the dependence of the network on hardware devices.Third,based on the densely connected network model,the lightweight Deep Lab V3+network is further improved,and a Deep Lab V3+ network with a densely connected structure is constructed.Two different structural improvement schemes are proposed to improve the segmentation performance of the network further.Fourth,to address the problems of inaccurate segmentation category recognition and inaccurate segmentation edges,an attention mechanism is introduced into the model,a segmentation model with a dual attention mechanism is constructed,and two different improvement schemes are designed with parallel and tandem structures.The proposed network model with a tandem attention mechanism is better optimized for segmentation performance.