Research On Segmentation And Annotation Method Of Remote Sensing Image Based On Scene Analysis

UAV remote sensing has the advantages of strong autonomy,simple operation,and fast response.It is widely used in many fields such as resource investigation,urban planning,and emergency handling.Its remote sensing image processing is the key technology and research focus.Aiming at the analysis of remote sensing object perception,this thesis carried out research on remote sensing image segmentation and semantic annotation methods based on scene analysis.The research results and innovations are as follows:1.Summarize the basic theoretical methods of remote sensing image segmentation and semantic annotation research,focusing on the traditional image segmentation methods based on clustering ideas,as well as the theoretical methods and related application cases of image classification and semantic annotation based on deep learning.2.Using the principle of Ncut image segmentation,an image segmentation method based on the combination of SLIC superpixels and graph theory Ncut is proposed.The SLIC superpixel algorithm is used to divide the image into multiple regions,and the graph theory Ncut method is used to calculate the similarity between the two superpixels to determine their attribute,and the superpixel regions are clustered;finally,the hierarchical region merge method is used to cluster the later regions are further merged,which can realize the region division of complex images.Experiments confirm that comparing with Ncut algorithm,this method reduces the segmentation time and also improves the segmentation effect.3.For the road extraction of remote sensing images,the improved road extraction method based on DeepLab V3+ is proposed,combined with the advantages of DeepLab V3+ network Atrous Spatial Pyramid Pooling(ASPP)and the advantages of U-Net fusion low level features and by combining DICE loss and BCE loss,this way can solve the problem of imbalance of binary classification samples,and the road extraction in Google map data set has better segmentation results.4.Aiming at the imbalance of samples in the deep neural network training in the labeling of multiple classes of features in remote sensing images,the ASPP module is used multiple times to fuse multi-scale features,the loss function uses weighted cross entropy loss,and is determined according to the proportion of samples in the training sample The weight can solve the problem that the model tends to fit the category with many samples.Through the experimental analysis of the CCF competition data set and the actual mining data set,the Multi-ASPP effect proposed in this paper is better than DeepLab V3+.5.Design a GUI display system that can perform human-computer interactive remote sensing image segmentation and semantic annotation.Call the remote sensing image segmentation method and the trained semantic segmentation model through the interface to test the image,and display the results on the interface;it can analyze the proportion of various classes of features in remote sensing images,and show the model's prediction accuracy and IOU;through human-computer interaction error correction and preservation processing,adding brushes to manually finetune annotation in the interface can achieve semi-automatic remote sensing image annotation.