CNNCounter.m

% Convolutional Neural Network Counter
% Authors: 
% Daniel Rodriguez
% Sebastian Salazar
% Reference: https://youtu.be/lK9YyX-q32k
close all
clear all
clc
%% Get images from folder
imgFolder = '/imgBinary';
imds = imageDatastore(imgFolder,...
    'IncludeSubFolders', true, 'LabelSource', 'foldernames');

%% Split data in 60% training, 20% validation and 20% test
fracTrainFiles = 0.6;
fracValFiles = 0.2;
fracTestFiles = 0.2;

[imdsTrain, imdsValidation, imdsTest] = splitEachLabel(imds, ...
    fracTrainFiles, fracValFiles, fracTestFiles, 'randomize');

layers = [
    imageInputLayer([150 150 1])
    
    convolution2dLayer(5,10,'Padding','same')
    batchNormalizationLayer
    reluLayer
    
    maxPooling2dLayer(2,'Stride',2)
    
    convolution2dLayer(5,10,'Padding','same')
    batchNormalizationLayer
    reluLayer
    
    fullyConnectedLayer(3)
    softmaxLayer
    classificationLayer
    ];

options = trainingOptions('sgdm', ...
    'InitialLearnRate',0.01, ...
    'MaxEpochs',100, ...
    'Shuffle','every-epoch', ...
    'ValidationData',imdsValidation, ...
    'ValidationFrequency',30, ...
    'Verbose',false, ...
    'Plots','training-progress');


net = trainNetwork(imdsTrain,layers,options);

%% Get accuracy
YPred = classify(net,imdsTest);
YTest = imdsTest.Labels;
accuracy = sum(YPred == YTest)/numel(YTest)

%% Show images data couldn't get predicted
ind = find(YPred ~= YTest);
figure; 
for ii = 1:length(ind)
    subplot(3,3,ii);
    imagesc(imdsValidation.readimage(ind(ii)));
    title([num2str(double(YPred(ind(ii)))-1), ' predicted, ', ... 
        num2str(double(YTest(ind(ii)))-1), ' actual'])
end