SRCNN代码分析

代码是作者页面上下载的matlab版。香港中文大学汤晓鸥教授。Learning a Deep Convolutional Network for Image Super-Resolution。

http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html

---

• demo_SR.m为运行主文件。

up_scale = 3;
model = 'model9-5-5(ImageNet)x3.mat';

• .mat文件存储数据。三个卷积层的权重和偏置。y=wx+b中的w和b。

%% work on illuminance only
if size(im,3)>1
    im = rgb2ycbcr(im);
    im = im(:, :, 1);
end
im_gnd = modcrop(im, up_scale);
im_gnd = single(im_gnd)/255;

彩色图像的话，RGB转为YCbCr。并只对Y通道处理。

若为灰度图像，直接处理。

其中，modcrop函数：将图片裁剪为能够调整的大小（与放大率匹配）。裁剪舍掉余数行和列。

function imgs = modcrop(imgs, modulo)
if size(imgs,3)==1
    sz = size(imgs);
    sz = sz - mod(sz, modulo);
    imgs = imgs(1:sz(1), 1:sz(2));
else
    tmpsz = size(imgs);
    sz = tmpsz(1:2);
    sz = sz - mod(sz, modulo);
    imgs = imgs(1:sz(1), 1:sz(2),:);
end

mod取余，crop修剪的意思。

double数据类型占8个字节，single类型占4个字节。Single（单精度浮点型）对图像归一化处理。得到im_gnd。

%% bicubic interpolation
im_l = imresize(im_gnd, 1/up_scale, 'bicubic');
im_b = imresize(im_l, up_scale, 'bicubic');

im_l ：将im_gnd 双三次插值缩小后的图像。

im_b : 将im_gnd 双三次插值缩小后再进行同比例放大的图像。

 

现在，LR是85*85的，放大三倍后是Bicubic，大小为255*255。

经过边界修剪，shave.m。四周各去边框-3。大小变为249*249。

function I = shave(I, border)
I = I(1+border(1):end-border(1), ...
      1+border(2):end-border(2), :, :);

或

function I = shave(I, border)
I = I(1+border(1):end-border(1),1+border(2):end-border(2));

结果是灰度图像：

想要生成彩色图像：可以见别人改过的代码。

【转载自】

SRCNN（一） - 刘一好 - 博客园 https://www.cnblogs.com/howtoloveyou/p/9691233.html

超分辨率重建SRCNN--Matlab 7.0中运行 - juebai123的博客 - CSDN博客 https://blog.csdn.net/juebai123/article/details/80532577 

SRCNN.m

function im_h = SRCNN(model, im_b)

%% load CNN model parameters
load(model);
[conv1_patchsize2,conv1_filters] = size(weights_conv1);
conv1_patchsize = sqrt(conv1_patchsize2);
[conv2_channels,conv2_patchsize2,conv2_filters] = size(weights_conv2);
conv2_patchsize = sqrt(conv2_patchsize2);
[conv3_channels,conv3_patchsize2] = size(weights_conv3);
conv3_patchsize = sqrt(conv3_patchsize2);
[hei, wid] = size(im_b);

%% conv1
weights_conv1 = reshape(weights_conv1, conv1_patchsize, conv1_patchsize, conv1_filters);
conv1_data = zeros(hei, wid, conv1_filters);
for i = 1 : conv1_filters
    conv1_data(:,:,i) = imfilter(im_b, weights_conv1(:,:,i), 'same', 'replicate');
    conv1_data(:,:,i) = max(conv1_data(:,:,i) + biases_conv1(i), 0);
end

%% conv2
conv2_data = zeros(hei, wid, conv2_filters);
for i = 1 : conv2_filters
    for j = 1 : conv2_channels
        conv2_subfilter = reshape(weights_conv2(j,:,i), conv2_patchsize, conv2_patchsize);
        conv2_data(:,:,i) = conv2_data(:,:,i) + imfilter(conv1_data(:,:,j), conv2_subfilter, 'same', 'replicate');
    end
    conv2_data(:,:,i) = max(conv2_data(:,:,i) + biases_conv2(i), 0);
end

%% conv3
conv3_data = zeros(hei, wid);
for i = 1 : conv3_channels
    conv3_subfilter = reshape(weights_conv3(i,:), conv3_patchsize, conv3_patchsize);
    conv3_data(:,:) = conv3_data(:,:) + imfilter(conv2_data(:,:,i), conv3_subfilter, 'same', 'replicate');
end

%% SRCNN reconstruction
im_h = conv3_data(:,:) + biases_conv3;

• l  Conv1: f1 = 9 *9 activation = ‘relu’
• l  Conv2: f2 = 1 *1 activation = ‘relu’  #为了非线性映射 增强非线性
• l  Conv3: f3 = 5 * 5 activation = ‘lienar’

【转载自】

SRCNN流程细节 - Python少年 - 博客园 https://www.cnblogs.com/echoboy/p/10289741.html