matlab 卷积神经网络 图像去噪 对抗样本修复

总流程类似于帮助文档帮助文档图像回归
神经网络结构类似于Unet，由卷积、激活、池化、转置卷积、深度连接等layers组成
先准备好数据集，我上传在资源里，可下载https://download.csdn.net/download/qq_51750957/14115738
1.matlab自带的0-9手写数字，共10000张，为[28,28]单通道图像

2.对应的对抗样本（加了干扰的图片）10000张（如何生成对抗样本可参照另一篇）

导入数据

imds = imageDatastore('DigitDataset', 'IncludeSubfolders', true, 'labelsource', 'foldernames');%读取原始图片
[imdsTrain,imdsValidation] = splitEachLabel(imds,0.9);%拆分出验证集
advs = imageDatastore('digit_adv0.15', 'IncludeSubfolders', true, 'labelsource', 'foldernames');%读取对抗样本
[advsTrain,advsValidation] = splitEachLabel(advs,0.9);

检查前十张图片是否对应

for i=1:10
    figure(i);
    subplot(1,2,1);imshow(imread(imds.Files{i}));
    subplot(1,2,2);imshow(imread(advs.Files{i}));
    list1{i}=imds.Files{i};
    list2{i}=advs.Files{i};
end

组合数据，形成输入输出

dsTrain = combine(advsTrain,imdsTrain);%合并为两列数据，第一列为对抗样本（输入），第二列为原始图片（输出）
dsVal = combine(advsValidation,imdsValidation);%同上
dsTrain = transform(dsTrain,@commonPreprocessing);%归一化到【0,1】,转为灰度图，缩放为[32,32]大小，commonPreprocessing定义在最后
dsVal = transform(dsVal,@commonPreprocessing);%同上

dsTrain=shuffle(dsTrain);%打乱

dataOut = readall(dsTrain);%显示前8对数据
inputs = dataOut(:,1);
responses = dataOut(:,2);
minibatch = cat(2,inputs,responses);
montage(minibatch','Size',[8 2])
title('Inputs (Left) and Responses (Right)')

设置训练参数

options = trainingOptions('adam', ...
    'MaxEpochs',30, ...
    'MiniBatchSize',32, ...
    'ValidationData',dsVal, ...
    'ValidationFrequency',300,...
     'ValidationPatience',100,...
    'Shuffle','never', ...
    'Plots','training-progress', ...
    'LearnRateSchedule','piecewise',...
    'LearnRateDropPeriod',5,...
    'LearnRateDropFactor',0.5,...
    'Verbose',false);

打开深度学习app，开始搭建网络，输入大小[32,32,1]，输出也是[32,32,1]，输入和输出每个像素都在【0,1】范围内。网络包含转置卷积层、深度连接层，比较复杂
搭建好后可导出为lgraph
网络很长，可以自己修改

创建深度学习网络架构
该脚本可创建具有以下属性的深度学习网络:
层数: 42
连接数: 44
运行脚本以在工作区变量 lgraph 中创建层。
要了解详细信息，请参阅 从 Deep Network Designer 生成 MATLAB 代码。
由 MATLAB 于 2021-01-11 20:52:38 自动生成
创建层次图
创建层次图变量以包含网络层。
lgraph = layerGraph();

添加层分支
将网络分支添加到层次图中。每个分支均为一个线性层组。
tempLayers = [
    imageInputLayer([32 32 1],"Name","imageinput","Normalization","rescale-zero-one")
    convolution2dLayer([3 3],32,"Name","conv_1_1","Padding","same")
    batchNormalizationLayer("Name","batchnorm_1")
    reluLayer("Name","relu_1_1")
    convolution2dLayer([3 3],32,"Name","conv_1_2","Padding","same")
    batchNormalizationLayer("Name","batchnorm_2")
    reluLayer("Name","relu_1_2")
    convolution2dLayer([3 3],32,"Name","conv_1_3","Padding","same")
    batchNormalizationLayer("Name","batchnorm_3")
    reluLayer("Name","relu_1_3")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    maxPooling2dLayer([2 2],"Name","maxpool_1","Padding","same","Stride",[2 2])
    convolution2dLayer([3 3],64,"Name","conv_2","Padding","same")
    reluLayer("Name","relu_2")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    dropoutLayer(0.15,"Name","dropout")
    maxPooling2dLayer([2 2],"Name","maxpool_2","Padding","same","Stride",[2 2])
    convolution2dLayer([3 3],128,"Name","conv_3","Padding","same")
    reluLayer("Name","relu_3")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    maxPooling2dLayer([2 2],"Name","maxpool_3","Padding","same","Stride",[2 2])
    convolution2dLayer([3 3],256,"Name","conv_5","Padding","same")
    batchNormalizationLayer("Name","batchnorm_6")
    reluLayer("Name","relu_7")
    transposedConv2dLayer([4 4],128,"Name","transposed-conv_1","Cropping",[1 1 1 1],"Stride",[2 2])
    reluLayer("Name","relu_4")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    depthConcatenationLayer(2,"Name","depthcat_3")
    convolution2dLayer([3 3],128,"Name","conv_1_4","Padding","same")
    batchNormalizationLayer("Name","batchnorm_4")
    reluLayer("Name","relu_1_4")
    transposedConv2dLayer([4 4],64,"Name","transposed-conv_2","Cropping",[1 1 1 1],"Stride",[2 2])
    reluLayer("Name","relu_5")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    depthConcatenationLayer(2,"Name","depthcat_1")
    convolution2dLayer([3 3],64,"Name","conv_1_5","Padding","same")
    batchNormalizationLayer("Name","batchnorm_5")
    reluLayer("Name","relu_1_5")
    transposedConv2dLayer([4 4],32,"Name","transposed-conv_3","Cropping",[1 1 1 1],"Stride",[2 2])
    reluLayer("Name","relu_6")];
lgraph = addLayers(lgraph,tempLayers);

tempLayers = [
    depthConcatenationLayer(2,"Name","depthcat_2")
    convolution2dLayer([3 3],32,"Name","conv_6","Padding","same")
    batchNormalizationLayer("Name","batchnorm_7")
    reluLayer("Name","relu_8")
    convolution2dLayer([1 1],1,"Name","conv_4","Padding","same")
    clippedReluLayer(1,"Name","clippedrelu")
    regressionLayer("Name","regressionoutput")];
lgraph = addLayers(lgraph,tempLayers);

% 清理辅助变量
clear tempLayers;

连接层分支
连接网络的所有分支以创建网络图。
lgraph = connectLayers(lgraph,"relu_1_3","maxpool_1");
lgraph = connectLayers(lgraph,"relu_1_3","depthcat_2/in2");
lgraph = connectLayers(lgraph,"relu_2","dropout");
lgraph = connectLayers(lgraph,"relu_2","depthcat_1/in1");
lgraph = connectLayers(lgraph,"relu_3","maxpool_3");
lgraph = connectLayers(lgraph,"relu_3","depthcat_3/in1");
lgraph = connectLayers(lgraph,"relu_4","depthcat_3/in2");
lgraph = connectLayers(lgraph,"relu_5","depthcat_1/in2");
lgraph = connectLayers(lgraph,"relu_6","depthcat_2/in1");

绘制层
plot(lgraph);

开始训练，训练耗时很长，可以适当更改maxEpoch

denoise_net = trainNetwork(dsTrain,lgraph,options);

训练完成，尝试预测5张图片，左边是输入，右边是输出

for ii=0:5
    figure(ii+1)
    I=imread(['digit_adv0.15/',num2str(ii),'/0000.png']);%读入图片
    I=imresize(I,[32,32]);
    %I=rgb2gray(I);%如果是彩图，要转为灰度
    I=single(I);
    I=rescale(I);
    subplot(1,2,1);imshow(I);
    ypred = predict(denoise_net,I);%预测
    subplot(1,2,2);imshow(rescale(ypred));
end

成功修复了简单的对抗样本

函数

function dataOut = commonPreprocessing(data)

dataOut = cell(size(data));
for col = 1:size(data,2)
    for idx = 1:size(data,1)
        temp = (data{idx,col});
        temp = imresize(temp,[32,32]);
        s=size(size(temp));
        if s(2)==3
        temp= rgb2gray(temp);
        end
        %temp = rescale(temp);
        dataOut{idx,col} = temp;
    end
end
end