稀疏表示字典的显示（MATLAB实现代码）

[置顶] 稀疏表示字典的显示（MATLAB实现代码）

2016-01-05 19:34 400人阅读 评论(0) 收藏 举报


分类：

ImageSuperResolution（10）

Compressive
Sensing（6）

Machine Learning（21）


版权声明：本文为博主原创文章，未经博主允许不得转载。

本文主要是实现论文--基于稀疏表示的图像超分辨率《Image Super-Resolution Via Sparse Representation》中的Figure2，通过对100000个高分辨率和低分辨率图像块训练得到的高分辨率图像块字典，字典原子总数为512，图像块尺寸大小为9X9



方法一：

[cpp] view
plain copy







clc;

clear all;

% load dictionary

load('Dictionary/D_512_0.15_9.mat');

patch_size=9;

nRow=24;

nCol=22;

D=Dh';

w=nCol*patch_size;

h=nRow*patch_size;

gridx = 1:patch_size :w;

gridx = [gridx, w-patch_size+1];

gridy = 1:patch_size : h;

gridy = [gridy, h-patch_size+1];

K=512; %字典原子总数

DD=cell(1,K);

row=length(gridx);

col=length(gridy);

hIm=zeros([w,h]);

for i=1:K

DD{i}=D(i,:);

end

for ii = 1:length(gridx),

for jj = 1:length(gridy),

yy = gridx(ii);

xx = gridy(jj);

if (ii-1)*nRow+jj >K

break

end

temp=DD{(ii-1)*nCol+jj};

hPatch=reshape(temp,[patch_size,patch_size]);

hIm(yy:yy+patch_size-1, xx:xx+patch_size-1) = hIm(yy:yy+patch_size-1, xx:xx+patch_size-1) +hPatch;

end

end

figure;

imagesc(hIm);

colormap(gray);

axis image;

输出结果：



可以看出，相比于论文中字典的显示图，颜色有点浅，通过调节MATLAB的colorbar，可以将背景颜色变深，

结果如下图所示：







方法二：

通过http://www.cs.technion.ac.il/~elad/software/提供的ksvd工具箱中的displayDictionaryElementsAsImage( )函数，来实现字典的显示。

displayDictionaryElementsAsImage.m

[cpp] view
plain copy







function I = displayDictionaryElementsAsImage(D, numRows, numCols,X,Y,sortVarFlag)

% function I = displayDictionaryElementsAsImage(D, numRows, numCols, X,Y)

% displays the dictionary atoms as blocks. For activation, the dictionary D

% should be given, as also the number of rows (numRows) and columns

% (numCols) for the atoms to be displayed. X and Y are the dimensions of

% each atom.

borderSize = 1;

columnScanFlag = 1;

strechEachVecFlag = 1;

showImFlag = 1;

if (length(who('X'))==0)

X = 8;

Y = 8;

end

if (length(who('sortVarFlag'))==0)

sortVarFlag = 1;

end

numElems = size(D,2);

if (length(who('numRows'))==0)

numRows = floor(sqrt(numElems));

numCols = numRows;

end

if (length(who('strechEachVecFlag'))==0)

strechEachVecFlag = 0;

end

if (length(who('showImFlag'))==0)

showImFlag = 1;

end

%%% sort the elements, if necessary.

%%% construct the image to display (I)

sizeForEachImage = sqrt(size(D,1))+borderSize;

I = zeros(sizeForEachImage*numRows+borderSize,sizeForEachImage*numCols+borderSize,3);

%%% fill all this image in blue

I(:,:,1) = 0; %min(min(D));

I(:,:,2) = 0; %min(min(D));

I(:,:,3) = 1; %max(max(D));

%%% now fill the image squares with the elements (in row scan or column

%%% scan).

if (strechEachVecFlag)

for counter = 1:size(D,2)

D(:,counter) = D(:,counter)-min(D(:,counter));

if (max(D(:,counter)))

D(:,counter) = D(:,counter)./max(D(:,counter));

end

end

end

if (sortVarFlag)

vars = var(D);

[V,indices] = sort(vars');

indices = fliplr(indices);

D = [D(:,1:sortVarFlag-1),D(:,indices+sortVarFlag-1)];

signs = sign(D(1,:));

signs(find(signs==0)) = 1;

D = D.*repmat(signs,size(D,1),1);

D = D(:,1:numRows*numCols);

end

counter=1;

for j = 1:numRows

for i = 1:numCols

% if (strechEachVecFlag)

% D(:,counter) = D(:,counter)-min(D(:,counter));

% D(:,counter) = D(:,counter)./max(D(:,counter));

% end

% if (columnScanFlag==1)

% I(borderSize+(i-1)*sizeForEachImage+1:i*sizeForEachImage,borderSize+(j-1)*sizeForEachImage+1:j*sizeForEachImage,1)=reshape(D(:,counter),8,8);

% I(borderSize+(i-1)*sizeForEachImage+1:i*sizeForEachImage,borderSize+(j-1)*sizeForEachImage+1:j*sizeForEachImage,2)=reshape(D(:,counter),8,8);

% I(borderSize+(i-1)*sizeForEachImage+1:i*sizeForEachImage,borderSize+(j-1)*sizeForEachImage+1:j*sizeForEachImage,3)=reshape(D(:,counter),8,8);

% else

% Go in Column Scan:

I(borderSize+(j-1)*sizeForEachImage+1:j*sizeForEachImage,borderSize+(i-1)*sizeForEachImage+1:i*sizeForEachImage,1)=reshape(D(:,counter),X,Y);

I(borderSize+(j-1)*sizeForEachImage+1:j*sizeForEachImage,borderSize+(i-1)*sizeForEachImage+1:i*sizeForEachImage,2)=reshape(D(:,counter),X,Y);

I(borderSize+(j-1)*sizeForEachImage+1:j*sizeForEachImage,borderSize+(i-1)*sizeForEachImage+1:i*sizeForEachImage,3)=reshape(D(:,counter),X,Y);

% end

counter = counter+1;

end

end

if (showImFlag)

I = I-min(min(min(I)));

I = I./max(max(max(I)));

imshow(I,[]);

end

测试程序

displayDictionary_test.m

[cpp] view
plain copy







clc;

clear all;

%加载字典

load('F:\Research\ScSR\ScSR\Dictionary\D_512_0.15_9.mat');

patch_size=9;

D=Dh;

K=512;

figure;

%调用KSVD工具箱中的字典显示函数

im=displayDictionaryElementsAsImage(D, floor(sqrt(K)), floor(size(D,2)/floor(sqrt(K))),patch_size,patch_size);

输出结果：





方法三：

因为方法一显示的字典图像偏灰，对比度不强，所以通过对字典原子像素值进行拉伸变化到0-1，增强图像对比度。

[cpp] view
plain copy







clc;

clear all;

% load dictionary

load('Dictionary/D_512_0.15_9.mat');

patch_size=9;

nRow=24;

nCol=22;

D=Dh';

w=nCol*patch_size;

h=nRow*patch_size;

gridx = 1:patch_size :w;

gridx = [gridx, w-patch_size+1];

gridy = 1:patch_size : h;

gridy = [gridy, h-patch_size+1];

K=512; %字典原子总数

DD=cell(1,K);

row=length(gridx);

col=length(gridy);

hIm=zeros([w,h]);

for i=1:K

DD{i}=D(i,:);

end

for ii = 1:length(gridx),

for jj = 1:length(gridy),

yy = gridx(ii);

xx = gridy(jj);

if (ii-1)*nRow+jj >K

break

end

temp=DD{(ii-1)*nCol+jj};

hPatch=reshape(temp,[patch_size,patch_size]);

I=hPatch;

I = I-min(min(min(I)));

I = I./max(max(max(I)));%对字典原子像素值进行拉伸变化到0-1

hIm(yy:yy+patch_size-1, xx:xx+patch_size-1) = hIm(yy:yy+patch_size-1, xx:xx+patch_size-1) +I;

end

end

figure;

imshow(hIm);



调整参数，将字典原子像素值拉伸变换到0-0.7

[cpp] view
plain copy







hPatch=reshape(temp,[patch_size,patch_size]);

I=hPatch;

I = I-min(min(min(I)));

I = 0.7*I./max(max(max(I)));%对字典原子像素值进行拉伸变化到0-0.7