The Classification Of SAR Images Based On Full Convolutional Network With Bayes Framework And Modification Of Explicit Edge Network

Synthetic Aperture Radar(SAR)has many advantages such as high resolution and multichannel compared to the traditional optical image.It is used to accomplish civil purpose such as crop observation and disaster observation,and also military purpose such as map protraction and military scout.Classically,we transform SAR signals into optical image,and apply methods used for optical image to classify the category of different objects or areas in SAR image.Our article use Pauli decomposition pseudo-image of ESAR,first we use traditional ways to classify and then apply methods of neural network and edge restriction,finally we raise the accuracy and improve the edge distribution compared to traditional methods.The classic way to classify SAR image need three main steps,including the segmentation of super pixel and feature extraction and classification machines accordingly.We often use meanshift or watershed methods to get super pixels,and apply(Support Vector Machine)SVM or(Back Propagation Network)BP to classification.Nowadays new methods such as(Markov Random Filed)MRF or(Conditional Random Field)CRF are introduced to improve the accuracy by taking nearby super pixel labels into account.First,tradition classifiers such as SVM,BP,Logistic classifier are designed to process 1-D signals instead of image,in order to improve the primary accuracy of classification,we introduce the(Fully Convolutional Network)FCN8 network which has the best performance in FCN series,and it can also conduct feature extraction,and meanwhile FCN directly compute on every single pixel instead of pixel group.Next,we introduce fully connected CRF in Deeplab to combine the self label probability with other pixel label around it as prior probability,and use grid optimization method to adjust the proper parameter.FCN has bad performance on edge area due to deconvolution step and convolutional kernel size.we can improve the classification result of edge area by(Domain Transform)DT network,but it has little effect.DT network has many critical problems,I analyze and solve the problems and propose a improved DT diffusion model.By generating edge map through(Holistic-nested Edge Detection)FCN-HED network,and combine the edge with the output of Deeplab,also with the output of edge network inside the DT can finally get a better result on accuracy and edge distribution,with the method we call as improved DT 4-D diffusion method and hole filling and inner fusion to get a explicit modification of classification.Last,we do lots of contrastive experiments on ESAR data,such as standard CRF based on potts or SVM classifier with meanshift to segmentation,FCN8,Deeplab and Deeplab-DT,and finally the experiment combining FCN-HED edge network with Deeplab-DT to explicit modification of edge classification effect.through the whole experiment we can conclude that the final experiment has the best performance on total classification accuracy and especially the improvement of edge area about both distribution and classification.