标签：tla dir training 处理 func 好的 影像 坐标 fprintf
近年来，深度学习在遥感影像地物分类中取得了一系列显著的效果。CNN可以很好的获取影像纹理信息，捕捉像素与像素之间的空间特征，因此，一个训练好的深度学习模型在地物提取中具有很大的优势。但模型的训练却是一个很繁琐的任务，需要人工准备数据集，贴标签，训练模型等。本文将以sar影像为例实现冰水二分类的数据集批量准备工作（划线取点截取小图片保存）：
1.原始sar遥感影像
2.预处理思路：
a.人工划线：对应在冰和水上画n条线（自己设置，注意自己需要针对类别所占比例控制线条数量和长度）
b.保存小图片：获取直线上点坐标，以每个像素点为中心取21×21的小图片（类似mnist数据集，尺寸自己设置），保存至文件夹
c. 创建label：以保存的小图片名称+空格+类别（0或者1）将label保存至新创建的txt文档中
3.代码实现：
a.创建一个main函数调用drawTrainingSamples(img);CreateTrainSmallImages(img);drawValSamples(img);CreateValSmallImages(img);这四个函数，功能分别是和划训练集，创建训练集，划验证集，创建验证集
clear ; clc; img = imread(‘150905_multilook_4_s1a-ew-grd-hv-20150905t174712-20150905t174812-007583-00a7f0-002.tiff‘); %准备训练集数据 drawTrainingSamples(img); CreateTrainSmallImages(img); %准备验证集数据 drawValSamples(img); CreateValSmallImages(img);
b.drawTrainingSamples(img)
function [] = drawTrainingSamples(img)
n_ice=4;
n_water=4;
h_im=imshow(img);
bw_train_ice=zeros(size(img));
bw_train_water=zeros(size(img));
fprintf(‘please draw four lines on the picture for preparing the training sets of Ice‘);
for i = 1:n_ice
h = imline;
bw = createMask(h,h_im);
bw_train_ice=bw_train_ice+bw;
end
figure,imshow(bw_train_ice);
h_im=imshow(img);
fprintf(‘please draw four lines on the picture for preparing the training sets of Water‘);
for i = 1:n_water
h = imline;
bw = createMask(h,h_im);
bw_train_water=bw_train_water+bw;
end
figure,imshow(bw_train_water);
save(‘bw_train_ice.mat‘,‘bw_train_ice‘);
save(‘bw_train_water.mat‘,‘bw_train_water‘);
c.CreateTrainSmallImages(img)
function [] = CreateTrainSmallImages(img)
%创建小图片
load bw_train_ice;
load bw_train_water;
fprintf(‘Creating training small images...‘);
[X,Y]=find(bw_train_ice==1);
A=[X,Y];
A;
[a,b]=size(A);
mkdir(‘train‘);
for i=1:a
m=A(i,1);
n=A(i,2);
SmallImage=img(m-10:m+10,n-10:n+10);
imwrite(SmallImage,[‘train/‘,num2str(i),‘.jpg‘]);
fid = fopen(‘train.txt‘, ‘a‘);
t=[num2str(i),‘.jpg‘];
fprintf(fid, ‘%s %d \n‘, t,0);
fclose(fid);
end
[X,Y]=find(bw_train_water==1);
B=[X,Y];
B;
[a,b]=size(B);
for j=1:a
m=B(j,1);
n=B(j,2);
SmallImage=img(m-10:m+10,n-10:n+10);
j=i+j;
imwrite(SmallImage,[‘train/‘,num2str(j),‘.jpg‘]);
fid = fopen(‘train.txt‘, ‘a‘);
t=[num2str(j),‘.jpg‘];
fprintf(fid, ‘%s %d \n‘, t,1);
fclose(fid);
end
end
d.drawValSamples(img)
function [] = drawValSamples(img)
n_ice=4;
n_water=4;
h_im=imshow(img);
bw_val_ice=zeros(size(img));
bw_val_water=zeros(size(img));
fprintf(‘please draw four lines on the picture for preparing the validition sets of Ice‘);
for i = 1:n_ice
h = imline;
bw = createMask(h,h_im);
bw_val_ice=bw_val_ice+bw;
end
figure,imshow(bw_val_ice);
h_im=imshow(img);
fprintf(‘please draw four lines on the picture for preparing the validition sets of Water‘);
for i = 1:n_water
h = imline;
bw = createMask(h,h_im);
bw_val_water=bw_val_water+bw;
end
figure,imshow(bw_val_water);
save(‘bw_val_ice.mat‘,‘bw_val_ice‘);
save(‘bw_val_water.mat‘,‘bw_val_water‘);
e.CreateValSmallImages(img)
function [] = CreateValSmallImages(img)
%创建小图片
load bw_val_ice;
load bw_val_water;
[X,Y]=find(bw_val_ice==1);
A=[X,Y];
A;
[a,b]=size(A);
mkdir(‘val‘);
fprintf(‘Creating validition sets small images...‘);
for i=1:a
m=A(i,1);
n=A(i,2);
SmallImage=img(m-10:m+10,n-10:n+10);
imwrite(SmallImage,[‘val/‘,num2str(i),‘.jpg‘]);
fid = fopen(‘val.txt‘, ‘a‘);
t=[num2str(i),‘.jpg‘];
fprintf(fid, ‘%s %d \n‘, t,0);
fclose(fid);
end
[X,Y]=find(bw_val_water==1);
B=[X,Y];
B;
[a,b]=size(B);
for j=1:a
m=B(j,1);
n=B(j,2);
SmallImage=img(m-10:m+10,n-10:n+10);
j=i+j;
imwrite(SmallImage,[‘val/‘,num2str(j),‘.jpg‘]);
fid = fopen(‘val.txt‘, ‘a‘);
t=[num2str(j),‘.jpg‘];
fprintf(fid, ‘%s %d \n‘, t,1);
fclose(fid);
end
end
标签：tla dir training 处理 func 好的 影像 坐标 fprintf
原文地址：http://www.cnblogs.com/xzyzg/p/7220823.html
