Handwriting recognition in MATLAB deep learning practice (including detailed notes)

Matlab In recent years, we have done better and better in artificial intelligence , Recently, in order to get familiar with matlab How to build neural networks , I did a handwriting recognition experiment , Make a note of .

The experimental task is very simple , The network is also very casual , Unreasonable places are also lazy to change , Designed to get through matlab The process of building neural network .

First , The data set is MNIST Data sets

I have sorted the data by category , It is divided into train and test, There are ten subfolders at the bottom to store handwritten images .

The network training code is as follows ：

clear;close all;clc;
%%  data fetch 、 enhance 
   % Read training set 
path_train = 'D:\work\ Expired files \ Handwriting Recognition \MNIST\train';  % Training set path 
folders_train = fullfile(path_train,{
    '0' '1' '2' '3' '4' '5' '6' '7' '8' '9'}); % Read subdirectories 
imds_train = imageDatastore(folders_train,'FileExtensions','.jpg',...
                            'LabelSource','foldernames');                 % Read all image paths 
[imdsTrain,imdsValidation] = splitEachLabel(imds_train,0.9,0.1);          % Split the validation set 
   % Read test set 
path_test = 'D:\work\ Expired files \ Handwriting Recognition \MNIST\test';
folders_test = fullfile(path_test,{
    '0' '1' '2' '3' '4' '5' '6' '7' '8' '9'});
imds_test = imageDatastore(folders_test,'FileExtensions','.jpg',...
                           'LabelSource','foldernames');
   % Image enhancement 
pixelRange = [-2 2];   % Translation range 
scaleRange = [0.9 1.1];  % Zoom range 
imageAugmenter = imageDataAugmenter( ...
    'RandXTranslation',pixelRange, ...
    'RandYTranslation',pixelRange, ...
    'RandXScale',scaleRange, ...
    'RandYScale',scaleRange);     % Define image intensifier 
augimdsTrain = augmentedImageDatastore([28,28],imds_train, ...
    'DataAugmentation',imageAugmenter);  % Image enhancement 
%%  Design （ Or read ） The Internet 
layers = [
    imageInputLayer([28 28 1],"Name","imageinput")
    convolution2dLayer([5 5],32,"Name","conv_1","Padding","same","Stride",[2 2])
    reluLayer("Name","relu_1")
    batchNormalizationLayer("Name","batchnorm_1")
    convolution2dLayer([3 3],32,"Name","conv_2","Padding","same")
    reluLayer("Name","relu_2")
    fullyConnectedLayer(512,"Name","fc_1")
    batchNormalizationLayer("Name","batchnorm_2")
    reluLayer("Name","relu_3")
    fullyConnectedLayer(10,"Name","fc_2")
    softmaxLayer("Name","softmax")
    classificationLayer("Name","classoutput")];
% analyzeNetwork(layers)  % Analyze the Internet 

%%  Training network 
options = trainingOptions('sgdm', ...
    'MiniBatchSize',512, ...
    'MaxEpochs',1, ...
    'InitialLearnRate',1e-2, ...
    'Shuffle','every-epoch', ...
    'ValidationData',imdsValidation, ...
    'ValidationFrequency',3, ...
    'Verbose',1, ...
    'Plots','training-progress');   % Set up training strategies 
trainedNet = trainNetwork(augimdsTrain,layers,options);  % Training 

%%  test model 
[YPred,probs] = classify(trainedNet,imds_test); 
accuracy = mean(YPred == imds_test.Labels)

Inside this , Some functions are used , I have written some important usage in other blogs , Here is just a general introduction to what is the use

• fullfile： Read all subfolders under the folder
• imageDatastore： Reading data sets , This function is more important , Many later functions are calling it
• splitEachLabel： Split imageDatastore Read data
• imageDataAugmenter： Image intensifier , Define how to enhance images
• augmentedImageDatastore： Image enhancement
• About How to build a network , I wrote it here , After reading it, I found that there is no need to type a line of code for some simple tasks
• analyzeNetwork： Analyze the Internet
• trainingOptions： Define training strategies , For example, learning rate , Optimizer and so on
• trainNetwork： Training network
• classify： Use the network for classification

Training results ：

If you need a processed data set , You can leave the mailbox ~

Listen to comments , Put it directly on the network disk , It should not be swallowed 2333
link ：https://pan.baidu.com/s/1htpPayQ2m0B3C5xOk3Quqg
Extraction code ：pdb6

Finally, let me explain , The network is built casually , Do not use ！！ Just learning MATLAB With

These hopes will help you