Research On Semantic Segmentation Of Remote Sensing Images Based On Self-attention Mechanis

Semantic segmentation of high-resolution remote sensing images plays a significant role in many practical applications,including precision agriculture and urban planning.With the development of deep learning,the semantic segmentation model based on convolutional neural network often has better segmentation performance,and has stronger real-time performance,so it is more suitable for practical situations.However,accurate segmentation of small-scale objects and edge details in remote sensing images is a huge challenge.With the deeper of the network model,low-level features containing geometric information and high-level features containing semantic information cannot be obtained at the same time,thus reducing the segmentation performance of the model.In addition,some semantic segmentation models based on convolutional networks do not fully refine the feature maps,and the context dependencies between feature maps have not been fully exploited,which weakens the expressive ability of features.In response to the above challenges,this paper summarizes the advantages and disadvantages of various model modules based on the deep network model,and proposes a high-precision and high-efficiency remote sensing image semantic segmentation model,and through sufficient experiments to verify the efficiency of the proposed model.The main contributions of this thesis are organized as follows:(1)We proposed a successive pooling attention network.The successive pooling attention module was proposed to organically couple the attention mechanism and multi-scale feature extraction to extract salient features in the image and suppress useless information,and complete the enhancement of the expressive ability of features.Then,a successive pooling strategy was adopted,while multi-scale features and deeper features can be extracted,which plays a very important role in the segmentation of small-scale objects and boundaries in remote sensing images.In addition,the feature fusion module realizes the complementarity of spatial information and geometric information,which alleviates the problem of information loss in the model.The efficiency of the proposed model performance was verified on the dataset.(2)We proposed a hierarchical residual refinement network to alleviate the bottleneck problem of low segmentation accuracy due to the lack of refinement of feature maps in the successive pooling attention network.The proposed channel attention module and pooled residual attention module fully mined the feature map position information or the dependence of information context between channels,thus enhancing the expressive ability of features.Then,the features extracted from the feature extractor at different stages and scales were fused step by step to realize the refinement of the feature map,and the fusion of multi-scale features also enhanced the model’s ability to recognize various types of ground objects,and the model’s generalization ability also was enhanced.In addition,by setting different residual structures,the correlation between gradient and loss during model training was improved,which enhanced the learning ability of the network and alleviated the problem of gradient disappearance.The efficiency of the proposed model performance is verified on the dataset.