Research On Semantic Segmentation Of SAR Image Based On U-net And Complex-Valued Segnet

Synthetic Aperture Radar(SAR)is an active earth observation system that actively emits energy and has the ability to work all-weather and all-day.It has been widely used in various fields such as marine glacier monitoring,earth resource survey,crop identification,etc.With the development of SAR systems,a large number of high-resolution SAR images have been produced,making SAR image interpretation technology particularly important.Among them,SAR image segmentation has always been one of the research hotspots in SAR image interpretation.Traditional SAR image segmentation methods often require manual feature extraction,which is time-consuming and labor-intensive,and the manually selected features are subjective.In recent years,with the development of deep learning,SAR image semantic segmentation has made certain progress.In the field of optical image semantic segmentation,the end-to-end semantic segmentation model has been deeply studied.Among them,some semantic segmentation models based on fully convolutional neural networks are widely used.Different from optical images,SAR images have the characteristics of coherent speckle noise and blurred edges of targets,which makes the semantic segmentation of SAR images more difficult.Although SAR image semantic segmentation based on deep learning has achieved certain results,there are still two more significant problems.First of all,deep learning requires a large number of training samples during training,and SAR image data is usually less,prone to over-fitting problems.Secondly,SAR data is complex data,and commonly used deep networks cannot make full use of the phase information in the complex data.In order to solve these problems,this paper proposes two SAR image semantic segmentation methods based on U-net and complex Segnet network respectively.The specific content is as follows:1)A SAR image semantic segmentation method based on U-net and Capsule Network(Capsule Network,Caps Net)is proposed.This method introduces a capsule network between the encoding and decoding parts of the U-net network to obtain the target's posture and other information.In addition,considering the small data set of SAR image,the U-net encoder is designed to be the same as VGG16,so that the trained VGG16 model can be directly transferred to the encoder.The experimental results of building target segmentation on three polarimetric SAR image data sets show that the proposed method based on U-net and capsule network in combination with transfer learning technology can not only obtain higher precision,recall,F1-score and intersection over union,but also can reduce the training time of network model in comparison with the method based on U-net.2)A method for semantic segmentation of SAR images based on complex-valued Segnet(CV-Segnet)is proposed.The CV-Segnet network is an extension of Segnet in the complex domain.It can make full use of the amplitude and phase information of the complex SAR image for target segmentation.The experimental results of image semantic segmentation on two polarized SAR data show that the complex valued Segnet has a higher pixel accuracy,mean pixel accuracy and mean intersection over union than real Segnet.