unsamplingmatlab代码的简单介绍

今天给各位分享unsamplingmatlab代码的知识，其中也会对进行解释，如果能碰巧解决你现在面临的问题，别忘了关注本站，现在开始吧！

本文目录一览：

• 1、matlab m文件代码能否生成模型
• 2、拉普拉斯金字塔图像融合的具体Matlab仿真程序
• 3、上采样(UnSampling)、反卷积(Deconvolution)与上池化(UnPooling)

matlab m文件代码能否生成模型

你是指用m文件写的数学模型，然后放在Simulink里面仿真是不？

如果是这样，Simulink里面有库中User-Defined

Functions/Embedded

MATLAB

Function

把这个块拖到新建模型中，双击打开，把你的m文件拷进去就行了。记得在m文件最前面加上

[y1,

y2

,

y3

...yn]

=

function(u1,u2,...un)

其中yi为你的模型的输出，ui为你模型的输入。

拉普拉斯金字塔图像融合的具体Matlab仿真程序

function lap_fusion()

%Laplacian Pyramid fusion

mul= imread('images\ms1.png');

pan= imread('images\pan.png');

figure(1);

imshow(mul);title('MS原始图像');axis fill;

figure(2);

imshow(pan);title('Pan原始图像');axis fill;

mul = double(rgb2gray(mul))/255;

pan = double(rgb2gray(pan))/255;

%普拉斯金塔变换参数

mp = 1;zt =4; cf =1;ar = 1; cc = [cf ar];

Y_lap = fuse_lap(mul,pan,zt,cc,mp);

figure(3);

imshow(Y_lap);title('lap fusion 后的图像');axis fill;

imwrite(Y_lap,'images\lap fusion后的图像.jpg','Quality',100);

%main function end

function Y = fuse_lap(M1, M2, zt, ap, mp)

%Y = fuse_lap(M1, M2, zt, ap, mp) image fusion with laplacian pyramid

%

% M1 - input image A

% M2 - input image B

% zt - maximum decomposition level

% ap - coefficient selection highpass (see selc.m)

% mp - coefficient selection base image (see selb.m)

%

% Y - fused image

% (Oliver Rockinger 16.08.99)

% check inputs

[z1 s1] = size(M1);

[z2 s2] = size(M2);

if (z1 ~= z2) | (s1 ~= s2)

error('Input images are not of same size');

end;

% define filter

w = [1 4 6 4 1] / 16;

% cells for selected images

E = cell(1,zt);

% loop over decomposition depth - analysis

for i1 = 1:zt

% calculate and store actual image size

[z s] = size(M1);

zl(i1) = z; sl(i1) = s;

% check if image expansion necessary

if (floor(z/2) ~= z/2), ew(1) = 1; else, ew(1) = 0; end;

if (floor(s/2) ~= s/2), ew(2) = 1; else, ew(2) = 0; end;

% perform expansion if necessary

if (any(ew))

M1 = adb(M1,ew);

M2 = adb(M2,ew);

end;

% perform filtering

G1 = conv2(conv2(es2(M1,2), w, 'valid'),w', 'valid');

G2 = conv2(conv2(es2(M2,2), w, 'valid'),w', 'valid');

% decimate, undecimate and interpolate

M1T = conv2(conv2(es2(undec2(dec2(G1)), 2), 2*w, 'valid'),2*w', 'valid');

M2T = conv2(conv2(es2(undec2(dec2(G2)), 2), 2*w, 'valid'),2*w', 'valid');

% select coefficients and store them

E(i1) = {selc(M1-M1T, M2-M2T, ap)};

% decimate

M1 = dec2(G1);

M2 = dec2(G2);

end;

% select base coefficients of last decompostion stage

M1 = selb(M1,M2,mp);

% loop over decomposition depth - synthesis

for i1 = zt:-1:1

% undecimate and interpolate

M1T = conv2(conv2(es2(undec2(M1), 2), 2*w, 'valid'), 2*w', 'valid');

% add coefficients

M1 = M1T + E{i1};

% select valid image region

M1 = M1(1:zl(i1),1:sl(i1));

end;

% copy image

Y = M1;

function Y = es2(X, n)

%Y = ES2(X, n) symmetric extension of a matrix on all borders

%

% X - input matrix

% n - number of rows/columns to extend

%

% Y - extended matrix

% (Oliver Rockinger 16.08.99)

[z s] = size(X);

Y = zeros(z+2*n, s+2*n);

Y(n+1:n+z,n:-1:1) = X(:,2:1:n+1);

Y(n+1:n+z,n+1:1:n+s) = X;

Y(n+1:n+z,n+s+1:1:s+2*n) = X(:,s-1:-1:s-n);

Y(n:-1:1,n+1:s+n) = X(2:1:n+1,:);

Y(n+z+1:1:z+2*n,n+1:s+n) = X(z-1:-1:z-n,:);

function Y = dec2(X);

%Y = dec2(X) downsampling of a matrix by 2

%

% X - input matrix

%

% Y - output matrix

% (Oliver Rockinger 16.08.99)

[a b] = size(X);

Y = X(1:2:a, 1:2:b);

function Y = undec2(X)

%Y = undec2(X) upsampling of a matrix by 2

%

% X - input matrix

%

% Y - output matrix

% (Oliver Rockinger 16.08.99)

[z s] = size(X);

Y = zeros(2*z, 2*s);

Y(1:2:2*z,1:2:2*s) = X;

function Y = selb(M1, M2, mp)

%Y = selb(M1, M2, mp) coefficient selection for base image

%

% M1 - coefficients A

% M2 - coefficients B

% mp - switch for selection type

% mp == 1: select A

% mp == 2: select B

% mp == 3: average A and B

%

% Y - combined coefficients

% (Oliver Rockinger 16.08.99)

switch (mp)

case 1, Y = M1;

case 2, Y = M2;

case 3, Y = (M1 + M2) / 2;

otherwise, error('unknown option');

end;

function Y = selc(M1, M2, ap)

%Y = selc(M1, M2, ap) coefficinet selection for highpass components

%

% M1 - coefficients A

% M2 - coefficients B

% mp - switch for selection type

% mp == 1: choose max(abs)

% mp == 2: salience / match measure with threshold == .75 (as proposed by Burt et al)

% mp == 3: choose max with consistency check (as proposed by Li et al)

% mp == 4: simple choose max

%

% Y - combined coefficients

% (Oliver Rockinger 16.08.99)

% check inputs

[z1 s1] = size(M1);

[z2 s2] = size(M2);

if (z1 ~= z2) | (s1 ~= s2)

error('Input images are not of same size');

end;

% switch to method

switch(ap(1))

case 1,

% choose max(abs)

mm = (abs(M1)) (abs(M2));

Y = (mm.*M1) + ((~mm).*M2);

case 2,

% Burts method

um = ap(2); th = .75;

% compute salience

S1 = conv2(es2(M1.*M1, floor(um/2)), ones(um), 'valid');

S2 = conv2(es2(M2.*M2, floor(um/2)), ones(um), 'valid');

% compute match

MA = conv2(es2(M1.*M2, floor(um/2)), ones(um), 'valid');

MA = 2 * MA ./ (S1 + S2 + eps);

% selection

m1 = MA th; m2 = S1 S2;

w1 = (0.5 - 0.5*(1-MA) / (1-th));

Y = (~m1) .* ((m2.*M1) + ((~m2).*M2));

Y = Y + (m1 .* ((m2.*M1.*(1-w1))+((m2).*M2.*w1) + ((~m2).*M2.*(1-w1))+((~m2).*M1.*w1)));

case 3,

% Lis method

um = ap(2);

% first step

A1 = ordfilt2(abs(es2(M1, floor(um/2))), um*um, ones(um));

A2 = ordfilt2(abs(es2(M2, floor(um/2))), um*um, ones(um));

% second step

mm = (conv2((A1 A2), ones(um), 'valid')) floor(um*um/2);

Y = (mm.*M1) + ((~mm).*M2);

case 4,

% simple choose max

mm = M1 M2;

Y = (mm.*M1) + ((~mm).*M2);

otherwise,

error('unkown option');

end;

上采样(UnSampling)、反卷积(Deconvolution)与上池化(UnPooling)

反卷积(Deconvolution)、上采样(UnSampling)与上池化(UnPooling)的对比图示如下（图片出处见水印）：

FCN、U-net等网络结构中都有使用上采样。上采样是指使图像分辨率变高的方法。

最简单的方式是重采样和插值：将输入图片进行rescale到一个想要的尺寸，并计算每个点的像素值，使用如 双线性插值 等插值方法对其余点进行插值来完成上采样过程。

反卷积在CNN中常用于表示一种反向卷积 ，但它并不是一个符合严格数学定义的反卷积操作。与上池化不同，使用反卷积来对图像进行上采样是可以习得的。通常用来对卷积层的结果进行上采样，使其回到原始图片的分辨率。

上池化是在CNN中常用的来表示max pooling的逆操作。最早来源于2013年纽约大学Matthew D. Zeiler和Rob Fergus发表的《Visualizing and Understanding Convolutional Networks》。

鉴于max pooling不可逆，因此使用近似的方式来反转得到max pooling操作之前的原始情况。

简单来说，记住做max pooling的时候的最大item的位置，比如一个3x3的矩阵，max pooling的size为2x2，stride为1，反卷积记住其位置，其余位置至为0就行：

unsamplingmatlab代码的介绍就聊到这里吧，感谢你花时间阅读本站内容，更多关于、unsamplingmatlab代码的信息别忘了在本站进行查找喔。