Research On SAR Image Road Extraction Method Based On Segmentation Network

By extracting road features and discriminating pavement elements,road segmentation in SAR images can realize accurate detection of road targets of different scales under different complex background conditions.It plays an important role in military and civilian fields such as battlefield monitoring,target positioning and tracking,map updating,geographic database construction and so on.The traditional method uses edge detection or region segmentation to extract the road,which has a lot of manual intervention and low efficiency and performance.Various methods based on complete convolutional neural network(FCN),such as U-Net,have excellent feature extraction ability and accurate segmentation effect,and are playing an increasingly important role in road extraction.However,compared with optical remote sensing images,SAR data sets often contain multiple resolution images with different road characteristics and interference.On the other hand,road extraction is faced with the problem of multiscale,and different receptive fields need to be combined,otherwise the extracted features are easy to produce omissions.Therefore,the current method is difficult to achieve satisfactory results.Aiming at the low segmentation accuracy of road targets in SAR image data with different resolutions,a multi-feature Cascade Unet(MFC-Unet)method based on dense join and Unet is proposed in this paper.By combining U-Net and dense connection,the method enhances the transmission and extraction of feature maps to reuse features more effectively,and cascades the multi-feature fusion of roads.In addition,considering the small proportion of road areas,this paper introduces a mixed attention module to guide the network to pay more attention to the road goal itself.Experimental results show that,compared with the existing advanced methods,the proposed method has advantages in different resolution SAR road extraction,especially for low and medium resolution SAR images,the integrity and accuracy of road extraction are greatly improved.For the problem that multi-scale roads in SAR images are difficult to be accurately divided simultaneously,a Dense Dilated Pyramid Unet(DDP-Unet)method based on void convolution is proposed in this paper.Based on FMC-Unet,this method uses void convolution to replace common convolution in dense connections,which can expand the dense sensitivity field of the network from different levels,and then extract rich details and multiscale information.It not only preserves the spatial features of the image,but also reduces the information loss without increasing the network parameter overhead.Experimental results show that the proposed method can extract SAR image roads more accurately,especially for multi-scale road images,avoiding the neglect of small roads.