Remote Sensing Image Segmentation With Multi-scale Deep Fusion Network And Attention Mechanism

With the rapid development of remote sensing technology,the form of remote sensing image data is more and more diversified,and the information contained in remote sensing image data is more and more abundant,which has important research value.As an important technology in remote sensing image processing,remote sensing image segmentation has been widely used in military reconnaissance,geological survey,map reconstruction and geological disaster prevention.In recent years,the remote sensing image segmentation methods based on deep learning have made some breakthroughs,but there are still some problems,such as the discontinuity of object edge segmentation and inaccurate prediction score of some objects.Aiming at these difficult problems,this thesis makes corresponding improvements from the network structure design to improve the accuracy of segmentation,and extends the research from two-dimensional image to three-dimensional point cloud.The main research contents of this thesis are as follows.1.Image instance segmentation method based on edge attention mechanism and deep evaluation network is proposed.Aiming at the problems of imprecise boundary box and incomplete edge segmentation in image instance segmentation task of mainstream network architecture,a full convolution deep evaluation network is designed to detect the object accurately and a supervised edge attention module is designed to enhance the edge features of instances.An evaluation branch is added to the full convolution deep evaluation network to learn association between the object feature and the detected bounding box.Finally,it provides high-quality bounding boxes for segmentation.By strengthening the edge features and suppressing the background noise,the edge attention module can get better edge segmentation effect,and then improve the accuracy of the instance segmentation.And the design of the network model structure is analyzed in detail,and the effectiveness of each module is verified in the experiment.The comparative experiments on iSAID dataset and MS COCO dataset all show the feasibility and robustness of the proposed method,and better segmentation results can be obtained by using this method.2.Image semantic segmentation method based on fine up-sampling and multi-scale dilated convolution network is proposed.The encoder-decoder network structure has been widely used in image semantic segmentation tasks,but the feature information will be lost in the down-sampling process.In order to solve this problem,multi-scale dilated convolution network is used to keep the spatial resolution of deep feature map,and multi-layer feature maps with different receptive field sizes are aggregated to obtain more comprehensive semantic feature information.In order to improve the accuracy of the image segmentation,a fine up-sampling structure network is used for further accurate segmentation of the uncertain pixels with lower score in the output of the segmentation network model.The comparative experiments are carried out on the US3 D dataset and the Cityscapes dataset,and the segmentation results of each model are visualized.Compared with some benchmark networks,the proposed method achieves the best performance on both datasets,which verifies the effectiveness and robustness of the proposed method.3.Multi-scale nested deep fusion network for 3D point cloud image segmentation method is proposed.In order to more fully extract and utilize the shallow and deep semantic features of the 3D point cloud,imitating the encoder-decoder network structure,a multi-scale nested deep fusion network based on PointNet++ is designed.The feature propagation module is inserted after each set abstraction module.In order to weaken the semantic gap,the feature maps with different levels of features are aggregated by using long and short skip link concatenation.A variable weight cross entropy loss function is designed to weaken the problem of extremely imbalanced data in 3D point cloud.The grid map method is used to correct the misclassification in the network model prediction results.It also gives a detailed analysis on the design of the network model structure.Finally,a series of comparative experiments are carried out on the large-scale urban remote sensing dataset US3 D,and ablation experiments are performed on the network structure.The experimental results show that the proposed method is better than the same type of methods.