Research And Implementation Of Remote Sensing Image Semantic Segmentation Based On Deep Learning

With the development of science and technology and the continuous progress of aerospace technology,massive remote sensing image data can be obtained,and the rapid and efficient processing of massive remote sensing data is the focus of current research.Remote sensing image semantic segmentation is the basis of remote sensing image processing,and it is also the focus of remote sensing image processing.However,in the semantic segmentation of remote sensing images,there are still problems such as difficulty in multi-scale target segmentation and blurred target edge segmentation.This thesis mainly studies the semantic segmentation method of remote sensing images that can obtain more scale features during feature extraction.Through densely connected atrous convolution,more feature information of different scales can be obtained without adding too many parameters,and the problem of multi-scale target segmentation difficulties in remote sensing image semantic segmentation can be solved.Aiming at the difficulty of multi-scale target segmentation in remote sensing images,a remote sensing image semantic segmentation model based on Densely connected Atrous Spatial Pyramid Pooling is proposed.The improved Densely connected Atrous Spatial Pyramid Pooling is used to enable the network to obtain more scale information and reduce The effect of small scale changes on the accuracy of semantic segmentation of remote sensing images.Aiming at the grid problem in the semantic segmentation results of the remote sensing image semantic segmentation model based on Densely connected Atrous Spatial Pyramid Pooling,Densely connected Atrous Spatial Pyramid Pooling module based on coprime factor is proposed to reduce the grid problem in the remote sensing image semantic segmentation results.grid problem.In order to enable the network to better obtain important features that are beneficial to improve the effect of semantic segmentation and ignore irrelevant parts,a remote sensing image semantic segmentation model combined with attention mechanism is proposed to improve the accuracy of remote sensing image semantic segmentation.Design and implement a remote sensing image semantic segmentation system,which can perform interactive remote sensing image semantic segmentation,call the trained remote sensing image semantic segmentation model through the interface to achieve remote sensing image semantic segmentation,and display the segmentation results on the interface.And it enables users to retrain the remote sensing image semantic segmentation model according to their own needs.In this thesis,the model is tested on the Deep Globe land cover classification dataset,and the performance of the model is evaluated by the precision and the Io U.The results show that the use of the Densely connected Atrous Spatial Pyramid Pooling module can improve the performance of remote sensing image semantic segmentation.The Densely connected Atrous Spatial Pyramid Pooling module based on coprime factor and the optimization of attention mechanism both achieve the effect of improving network performance.