Research On Remote Sensing Image Semantic Segmentation Algorithm Based On Deep Learning

With the development of aviation and space technology technology,it is more and more convenient for people to obtain remote sensing images.Nowadays,remote sensing image semantic segmentation technology is a popular research direction,which has important significance in urban construction,land cover mapping,urban planning,military security,environmental protection and fine agriculture.Many excellent semantic segmentation methods have been proposed and applied in the field of remote sensing.In recent years,improving the accuracy of remote sensing image semantic segmentation based on deep learning has become a common research goal among scholars.This article uses deep learning methods to study the semantic segmentation technology of remote sensing images,and proposes two semantic segmentation algorithms(RTNet and EHRTNet networks).Experiments are conducted on the ISPRS Vaihingen dataset and Potsdam dataset,and excellent segmentation results are achieved.The work content of this article is as follows:Firstly,based on the Deep Labv3+network,a segmentation network RTNet based on global feature optimization is constructed for semantic segmentation of remote sensing images,achieving more comprehensive global feature extraction in the field of high-resolution remote sensing images.(1)We have used the latest Deep Labv3+segmentation framework for semantic segmentation of remote sensing images,and proposed the ASPPBT module.When extracting global features,we can create a close relationship between the global features extracted by the network.(2)By replacing the original Res Net skeleton network with the latest Rep VGG network structure,we found that the Rep VGG skeleton network is more suitable for remote sensing image segmentation compared to Res Net,and the effect will be better.(3)Experiments have shown that our proposed RTNet remote sensing image semantic segmentation network has a more robust effect on the Vaihingen dataset.Even under the influence of cloud cover and complex surrounding terrain,this architecture can effectively separate various objects in remote sensing images.Compared with other advanced segmentation algorithms in the field of remote sensing,our architecture can segment them more accurately.Meanwhile,the generalization of the RTNet network was verified on the Potsdam dataset.This can demonstrate the effectiveness of our method.Secondly,based on the above research,a segmentation network EH-RTNet based on edge guidance and gradient coordination strategy is proposed for semantic segmentation of remote sensing images to address the issues of edge guidance and sample balance training.This achieves the goal of edge protection in the field of high-resolution remote sensing images.(1)Based on the above research,a multi-scale edge detection EDMS module is proposed,which can guide the network to better perceive object edges during the training process.(2)By replacing the original loss function with the latest gradient coordination mechanism GHM,we find that GHM-C is more suitable for remote sensing image training than the original loss function,and can improve the training efficiency and detection accuracy of the model.(3)The experiment shows that the segmentation results of our proposed EH-RTNet remote sensing image semantic segmentation network are more precise and can be well applied to remote sensing image semantic segmentation.