1、相關(guān)理論

CNN 模型為深度學(xué)習(xí)模型，其具有局部連接、權(quán) 值共享和空間相關(guān)等特性，以及強(qiáng)魯棒性和容錯(cuò)能力，適用于提取深層數(shù)據(jù)特征。經(jīng)典的 CNN 模型結(jié)構(gòu)包含輸入層、隱含層、全連接層和輸出層。卷積神經(jīng)網(wǎng)絡(luò)的模型如下圖所示。

2、數(shù)據(jù)集的準(zhǔn)備

以手寫數(shù)據(jù)集為例，搭建卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行分類識(shí)別。數(shù)據(jù)集下載地址： ，下面為部分?jǐn)?shù)據(jù)集圖片展示。

3、數(shù)據(jù)集的讀取與劃分

將下載好的數(shù)據(jù)集保存好，digitDatasetPath 填寫數(shù)據(jù)集的保存路徑即可。每一類隨機(jī)選擇750張圖片作為測(cè)試數(shù)據(jù)，其他的作為訓(xùn)練數(shù)據(jù)。

%%數(shù)據(jù)集的讀取
digitDatasetPath = 'D:\\MTALAB2019\\手寫數(shù)據(jù)集\\DigitDataset';
imds = imageDatastore(digitDatasetPath, ...
    'IncludeSubfolders',true,'LabelSource','foldernames');
%%數(shù)據(jù)集的劃分
numTrainFiles = 750;
[imdsTrain,imdsValidation] = splitEachLabel(imds,numTrainFiles,'randomize');

4、卷積神經(jīng)網(wǎng)絡(luò)的搭建

layers = [
    imageInputLayer([28 28 1])  %%輸入層
    %%卷積層
    convolution2dLayer(3,8,'Padding','same')  
    batchNormalizationLayer
    reluLayer
    %%池化層
    maxPooling2dLayer(2,'Stride',2)
    %%卷積層
    convolution2dLayer(3,16,'Padding','same')
    batchNormalizationLayer
    reluLayer
    %%池化層
    maxPooling2dLayer(2,'Stride',2)
    %%卷積層
    convolution2dLayer(3,32,'Padding','same')
    batchNormalizationLayer
    reluLayer
    %全連接層
    fullyConnectedLayer(10)
    softmaxLayer
    classificationLayer];

網(wǎng)絡(luò)搭建好了，就需要對(duì)網(wǎng)絡(luò)的參數(shù)進(jìn)行設(shè)置，相關(guān)參數(shù)代碼如下：

options = trainingOptions('sgdm', ...
    'InitialLearnRate',0.01, ...
    'MaxEpochs',10, ...
    'Shuffle','every-epoch', ...
    'ValidationData',imdsValidation, ...
    'ValidationFrequency',30, ...
    'Verbose',false, ...
    'Plots','training-progress');

** 5、訓(xùn)練卷積神經(jīng)網(wǎng)絡(luò)**

net = trainNetwork(imdsTrain,layers,options);

** 訓(xùn)練結(jié)果如下 **

6、測(cè)試與運(yùn)行結(jié)果

YPred = classify(net,imdsValidation);
YValidation = imdsValidation.Labels;
accuracy = sum(YPred == YValidation)/numel(YValidation)

accuracy = 0.9868