1 简介
DPPCANet是一种鲁棒的深度学习方法，用于不平衡多时相合成孔径雷达图像的变化检测，主要包括1）生成差异图；2）并行FCM聚类，提供训练样本伪标签；3）基于采样PCANet+SVM模型构建过采样和欠采样像素分类。
2 部分代码
% notice：% The training samples are randomly generated, so the results are slightly different% clear;% clc;% close all;addpath('./utils');addpath('./liblinear');pool_size = 3;w = 7;b = 0.16;T_num = 11;% Import Dataim1 = imread('./pic/B1.tif');im2 = imread('./pic/B2.tif');im_gt = imread('./pic/BGT.tif');im1 = double(im1(:,:,1)); im2 = double(im2(:,:,1)); im_gt = double(im_gt(:,:,1));% Compute the deep difference imagefprintf('... ... compute the deep difference image ... ...\n');im1 = WP(im1,pool_size);im2 = WP(im2,pool_size);DI_or = di_gen(im1,im2); % Calculate log-ratio imageDDI = Normalized(CWP(DI_or,T_num)); % Calculate deep different image[DDIMAP1,DDIMAP2]=DDIMAP(DI_or,T_num,w,b); % Computing mapped DDI% Gabor feature extractionfprintf('... ... Gabor feature extraction... ...\n');[f1_all,fea_1] = Gabor_fea(DDIMAP1);[f2_all,fea_2] = Gabor_fea(DDIMAP2);%Parallel FCM clusteringfprintf('... ... parallelclustering begin ... ...\n');im_lab = parallelclustering(fea_1,DDIMAP1,fea_2,DDIMAP2);% Clustering results are saved in im_lab% Changed pixels as 1% Unchanged pixels is marked as 0% Classification modelPatSize = 5; % PatSize must be oddim_lab = 1-im_lab;% PCANet PCANet_SVM_train;PCANet_SVM_test;[Ylen, Xlen] = size(im_gt);% DenoisingPreRes = reshape(PreRes, Ylen, Xlen);[lab_pre,num] = bwlabel(~PreRes);for i = 1:num idx = find(lab_pre==i); if numel(idx) <= 20 lab_pre(idx)=0; endend%Save resultslab_pre = lab_pre>0;res = uint8(lab_pre)*255;pic = res;[TP,TN,FP,FN,MC,MU,FPR,FNR,OER,PCC,Kappa] = PE(res,im_gt);list = [TP,TN,FP,FN,MC,MU,FPR,FNR,OER,PCC,Kappa];imwrite(pic, 'changemap.png');save('result.mat','list');figureim1 = imread('./pic/B1.tif');im2 = imread('./pic/B2.tif');subplot(131);imshow(im1);title('图1')subplot(132);imshow(im2);title('图2')subplot(133);imshow(pic);title('检测图')
