upload

demo.m

@@ -0,0 +1,29 @@
+%% ³õʼ»¯
+
+clc;
+clear;
+addpath('./src')
+
+%% ÔØÈë²âÊÔͼÏñ
+
+im = imread('./img/100080.jpg');
+im = im(201:340, 101:240, 1);
+[H, W, ~] = size(im);
+input = reshape(im, 1, 1, H, W);
+
+%% run
+
+tic
+upscale_factor = 5;
+model_name = ['./model/model_upscale_', num2str(upscale_factor),'.mat'];
+im_sr = SR_model(input, model_name);
+size(im_sr)
+toc
+%% ÏÔʾ
+
+figure
+[~,~,h_sr, w_sr] = size(im_sr);
+imshow(reshape(im_sr, h_sr, w_sr), [])
+figure
+[~,~,h, w] = size(input);
+imshow(reshape(input, h, w), [])

src/PixelShuffle.m

@@ -0,0 +1,22 @@
+function [ outputs ] = PixelShuffle( inputs, upscale_factor )
+%	PixelShuffle :
+%
+%   input : N, upscale_factor ** 2, H, W
+%   output : N, 1, H*upscale_factor, W*upscale_factor
+
+[N, ~, H, W] = size(inputs);
+H_out = H*upscale_factor;
+W_out = W*upscale_factor;
+outputs = zeros([N, 1, H_out, W_out]);
+for i = 1:N
+    for h = 1: H_out
+        for w = 1:W_out
+            height_idx = floor(h / upscale_factor+0.5);
+            weight_idx = floor(w / upscale_factor+0.5);
+            channel_idx = (upscale_factor * mod(h-1, upscale_factor)) + mod(w-1, upscale_factor)+1;
+            outputs(i, 1, h, w) = inputs(i, channel_idx, height_idx, weight_idx);
+        end
+    end
+end
+end
+

src/SR_model.m

@@ -0,0 +1,40 @@
+function [ image_sr ] = SR_model( image, model_name)
+%{ 
+一、函数概述
+    SR_model, 该函数以单波段图像作为输入，使用卷积神经网络做超分辨处理，输
+出超分辨后的图像。不同的超分辨因子对应不同的预训练模型，超分辨因子可指定
+2-10中的任意整数,指定超分辨因子时，手动指定不同的权重即可。权重命名具有一
+定规律，如 model_upscale_7.mat，表示超分辨因子为7时对应的模型。
+    该函数使用高效的子像素卷积层进行图像超分辨[1], 在GPU加速条件下能够实现
+720P视频实时超分辨，相比于之前的深度学习模型，推理速度得到了极大提高。
+
+二、参数说明：
+  image ：4维数组，（N, 1, H, W）,N 为图像数量
+  model : 预训练的权值
+  image_sr ：4维数组，超分辨后的图像序列， 
+  尺寸为：（N, 1, H * upscale_factor, W * upscale_factor）
+  其中 N 为图像数量
+
+三、参考文献：
+[1] Shi W, Caballero J, Huszar F, et al. Real-Time Single Image and 
+    Video Super-Resolution Using an Efficient Sub-Pixel Convolutional 
+    Neural Network[J]. 2016:1874-1883.
+%}
+
+load(model_name)
+out1 = conv_func(image, conv1_w);
+out1 = relu(out1 + conv1_b);
+
+out2 = conv_func(out1, conv2_w);
+out2 = relu(out2 + conv2_b);
+
+out3 = conv_func(out2, conv3_w);
+out3 = relu(out3 + conv3_b);
+
+out4 = conv_func(out3, conv4_w);
+out4 = out4 + conv4_b;
+
+image_sr = PixelShuffle( out4, 2 );
+
+end
+

src/conv_func.m

@@ -0,0 +1,28 @@
+function [ outputs ] = conv_func( inputs, kernels )
+%	conv_func : 深度学习中的卷积层前向运算
+%   with stride = 1 and padding = 'same'
+%
+%   input : N, C_in, H, W
+%   kernel : C_out, C_in, kernel_h, kernel_w
+%   output : N, C_out, H, W
+inputs = double(inputs);
+kernels = double(kernels);
+[N, ~, H, W] = size(inputs);
+[C_out, C_in, kernel_h, kernel_w] = size(kernels);
+
+outputs = zeros(N, C_out, H, W);
+tmp = zeros(C_in, H, W);
+
+for i = 1:N
+    for j = 1:C_out
+        for k = 1:C_in
+            input = reshape(inputs(i, k, :, :), H, W);
+            kernel = reshape(kernels(j, k, :, :), kernel_h, kernel_w);
+            tmp(k, :, :) = filter2(kernel, input, 'same'); % tmp (H, W)
+            outputs(i, j, :, :) = sum(tmp, 1);
+        end
+    end
+end
+
+end
+

src/relu.m

@@ -0,0 +1,8 @@
+function [ output ] = relu( input )
+% Éñ¾­ÍøÂçÖеÄReLU¼¤»îº¯Êý
+% ReLU(x) = max(x, 0)
+input (input < 0) = 0;
+output = input;
+
+end
+