Research On Low Scattering Region Segmentation Of SAR Image

As an advanced earth observation tool,synthetic aperture radar(SAR)has high application value in military and civil fields.SAR image segmentation is an important technology of SAR image interpretation,but there is little research on low scattering region segmentation at home and abroad.The low scattering area of SAR image studied in this paper includes shadow,water area,road and airport runway.The research on their segmentation is of great practical significance in three aspects: environment and flood disaster monitoring,military target strike and UAV landing,and three-dimensional reconstruction of surface objects.They are dark regions in SAR images,so it is difficult to segment them.Therefore,this paper studies the data set making and segmentation technology of SAR image low scattering region.Firstly,this paper solves the problem of making SAR image data set with low scattering area,and adopts a label annotation method based on Photoshop,Labelme and Python.It is a very time-consuming problem to label segmented data labels manually.Our method greatly improves the labeling efficiency and has high accuracy in labeling low scattering areas of SAR images,and it is more precise for some complex edges than using Labelme only.Secondly,the multi module fusion network(MMFNet)is proposed to segment the low scattering region of SAR image.MMFNet is composed of high-resolution backbone network module,spatial pyramid pooling convolution module and channel attention module.The high-resolution backbone network can maintain the high-resolution characteristics of the feature map,reduce the loss of spatial accuracy,it is helpful to extract edge information and other detail information and promote the detection of small target.The spatial pyramid pooling convolution module can fuse multi-scale feature,which is conducive to the extraction of different sizes of targets,and improve the accuracy of segmentation.The channel attention module can enhance the network's ability to express category information and improve the accuracy of category segmentation.In this paper,the segmentation effect of MMFNet is verified by experiments on SAR image low scattering region data set.The classical high-precision semantic segmentation networks Deep Lab V3+,DANet,Enc Net and HRNet + OCR are used to do the comparative experiments.The results show that MMFNet achieves good segmentation effect,and the segmentation accuracy is better than Deep Lab V3+,DANet,Enc Net and HRNet + OCR.Finally,the codec network(CNet)is proposed for fast segmentation of low scattering regions in SAR images.CNet adopts asymmetric encoding and decoding structure.The encoder can extract features of different levels from shallow to deep,while the decoder fuses the features of different levels and outputs them at last,and maintaining the high resolution of the feature map.In the early stage of coding,downsampling is used to reduce the subsequent computation of the model and improve the segmentation speed.The decoder improves the segmentation speed by compressing the network volume and reducing the complex convolution operation.We use experiments to verify CNet,and use the classic real-time semantic segmentation network Seg Net,ENet and BiSeNet to do comparative experiments.Experimental results show that our proposed CNet not only has fast segmentation speed,but also has high segmentation accuracy.And CNet is better than Seg Net,ENet and BiSeNet in speed and accuracy.