Research And Application Of Semantic Segmentation Algorithm For Remote Sensing Images

The current convolutional neural network serves as the dominant model for the se-mantic segmentation of remote sensing images.However,owing to the inherent nature of convolution,convolutional neural networks suffer from limited perceptual fields and fail to provide sufficient contextual information.To obtain multi-level and wide-range contex-tual information as possible,existing semantic segmentation methods implicitly expand the perceptual field by downsampling the spatial resolution of the feature map.However,the downsampling operation has limited ability to expand the perceptual field and leads to the extra problem of missing spatial detail information,which is one of the key factors to determine the performance of semantic segmentation.Therefore,the decoder of the semantic segmentation algorithm needs to restore the low-resolution high-level semantic feature maps to high-resolution classification maps.The existing methods generally use bilinear upsampling for interpolation calculation.Nevertheless,the feature map detail in-formation after upsampling is only calculated by interpolation,which cannot accurately describe the missing actual detail information,which is reflected in the blurred and in-accurate position of the object boundary on the segmentation result.In addition,the gap between the current cutting-edge algorithms of semantic segmentation of remote sensing images and real application scenarios has not been bridged,which lacks out-of-the-box applications.In this paper,we attempt to propose solutions for the above problems from different perspectives.The main work of this paper is as follows:1.A method for modelling attention in different level contextual scopes.This method improves the existing attention mechanism which lacks local detail information,is computationally expensive and has unclear attention formation.By explicitly modelling attention at spatial scope at three levels: local level,semantic level and image level,the explicitness of attention is improved while effectively reducing computation and memory usage,and local level attention also improves the problem of lack of detail information in decoder feature maps.2.A region classification-based image segmentation method is proposed.Different from the segmentation method of pixel-level classification,this method breaks down the semantic segmentation task into two subtasks,region segmentation and region classification.The region segmentation task segments the image into several re-gions which have clear geometric shapes and semantic consistency.The region classification task predicts the categories of all regions,and the semantic segmen-tation results are obtained by filling the region classification results into the cor-responding regions.This method effectively solves the problem of blurred object boundaries and inaccurate localization in segmentation results with the region ge-ometry information predicted by the region segmentation branch and improves the problem of insufficient contextual information with the self-attention mechanism in the region classification branch.3.A remote sensing image segmentation system is designed based on the two meth-ods mentioned previously.The system closes the gap between the current state-of-the-art technology of semantic segmentation of remote sensing images and real application scenarios,and provides out-of-the-box solutions for application scenar-ios such as smart city construction and monitoring of land resources.The system provides four types of interpretation tasks: building extraction,grass extraction,tree extraction and vehicle extraction.In terms of design architecture,the system adopts separate front-end and back-end architecture,where the front-end can be deployed on portable devices such as cell phones and laptops,and interacts with the back-end server set up at remote locations through network communication.