Multi-scale Sparse Fusion Network Based Remote Sensing Image Segmentation

Remote sensing image is widely used in traditional geography,environmental science and geoscience fields because of its low cost,high imaging accuracy,continuous monitoring of large-scale land,and not limited by weather and environmental conditions.It is of great significance for geographic exploration,people travel and national security.In recent years,with the development of imaging methods and aerospace technology,a large number of high-resolution remote sensing images have been obtained,and the importance of remote sensing images understanding and translation is increasing day by day.As an important basis in this field,the remote sensing image segmentation has been widely concerned by many scholars.Semantic segmentation of remote sensing image is a key technology in remote sensing image processing.It is of great significance to further research on remote sensing image interpretation,and has become a hot issue in recent years.Compared with ordinary urban or indoor scenes,remote sensing image semantic segmentation is difficult because of the rich semantic information,various features,and different sizes in remote sensing images,which puts forward higher requirements on semantic segmentation networks.This thesis aims at the characteristics of remote sensing image and combines the popular deep learning network,and makes a deep study and discussion on the semantic segmentation of remote sensing images.The main contributions of this thesis are as follows.(1)Considering the rich semantic information and the different sizes of objects in the remote sensing images,a segmentation method of remote sensing image based on spatial information and multi-scale fusion is proposed.Seg Net is used as the backbone network in this method,and improves the recognition accuracy of objects of different sizes by adding skip connection structure and spatial information fusion module.Secondly,due to the large scale of remote sensing images,they need to be cropped before they can be processed,so the problem of blocking is prone to occur.This thesis also proposes an effective smoothing strategy to improve the quality of image segmentation.Experiments show that the proposed semantic segmentation method of large-scale remote sensing image based on spatial information and multi-scale fusion achieves accurate segmentation results.(2)Since complex types and shadow occlusion of remote sensing images,a remote sensing image segmentation method based on sparse space fusion network is proposed.Sparse representation is introduced into the deep network to improve the performance of feature discrimination and reduce the adverse effects of shadows on the segmentation results.Fully connected layer without bias term and nonlinear activation is used to achieve sparse self-representation,which is added to the proposed network to obtain a new loss function with sparse regularization term.Finally,the segmentation results on ISPRS data set and remote sensing image data set prove that sparse self-representation can effectively improve the performance of the network.(3)In order to make full use of the contextual semantic information of images,uses spatial pathways and densely connected blocks to further improve the performance of network.Context information is very important for semantic segmentation task,but the network with too many layers will lead to the loss of features,which has a great impact on the semantic segmentation task that needs intensive prediction.This section introduces spatial pathways to enhance the network’s learning of the underlying information,and at the same time uses densely connected blocks,so that higher accuracy can be achieved at a small computational cost without changing the original network framework.Finally,experiments prove the effectiveness of the proposed method.