Research And Application Of An Edge-Aware Two-Branch Semantic Segmentation Model For High Resolution Remote Sensing Images

With the rapid development of aerospace remote sensing technology and sensor technology,the quality and quantity of remote sensing information obtained through remote sensing technology have also made a qualitative leap,among which the generation of high resolution remote sensing images provides the conditions and basis for effective geo-interpretation and analysis.Image segmentation is the key step of remote sensing image interpretation,and its segmentation result directly determines the accuracy of subsequent interpretation.Compared with other types of remote sensing images,high resolution remote sensing images contain richer feature information and are characterized by complex backgrounds and diverse targets,and traditional image segmentation methods can no longer meet the segmentation accuracy requirements of existing tasks.In recent years,deep learning,especially convolutional neural networks,has been gradually applied in the field of image semantic segmentation,and also promoted the development of remote sensing image semantic segmentation.This paper focuses on the application of current semantic segmentation algorithms based on convolutional neural networks in high resolution remote sensing images,and the main research contents are as follows.Aiming at the current problems of large data volume and inability to directly input segmentation when convolutional neural networks are applied to semantic segmentation of high-resolution remote sensing images,a light-weight semantic segmentation network model with global and local two-branch network structure is proposed.The model contains a global path and a local path,where the input of the global path is the global image that has been downsampled and the input of the local path is the local image with poor initial segmentation effect.On the one hand,the rich spatial structure information of the global view is used to resolve the ambiguity of the local segmentation,and on the other hand,the local segmentation is used to refine the details lost by the global segmentation due to downsampling.The experimental results on high resolution remote sensing datasets show that the network model proposed in this paper achieves a balance in accuracy and efficiency compared with other segmentation models.To address the problem of inaccurate feature boundary segmentation when current semantic segmentation models are used for high resolution remote sensing images,an enhanced semantic segmentation model based on edge perception is proposed,whose coding structure contains two paths,edge path and spatial path,and edge prediction is used as an independent self-task to assist the segmentation model,while an edge-guided contextual aggregation module is designed to effectively aggregate edge features and semantic features to achieve the segmentation effect of edge enhancement.We validate and prove the effectiveness of our method on a high resolution remote sensing image dataset.Finally,we design a semantic segmentation system based on high resolution remote sensing images,which can perform semantic segmentation on user-uploaded high-resolution remote sensing images and support single-image segmentation,multi-image segmentation and multi-algorithm segmentation comparison,which can provide diversified services for users.