Main.m

% Deep Learning Assignment Project 5
% Brian Zenger, U0291777
% Bioen 6640 Image Processing

%% run this during the matlab startup if networks already created
vl_setupnn;
vl_compilenn;

tempnet = load('matconvnet-1.0-beta25/data/mnist-baseline-simplenn/net-epoch-20.mat')
net = tempnet.net
tempnet = load('matconvnet-1.0-beta25/catdog/net-epoch-80.mat')
netC = tempnet.net
tempnet = load('matconvnet-1.0-beta25/data/mnist-evenodd-simplenn/net-epoch-20.mat')
netEO = tempnet.net
%% Viewing the weights of the first layer
%montage(net.layers{1}.weights{1})

figure(1)
for k = 1:20
subplot(4,5,k)
imagesc(net.layers{1}.weights{1}(:,:,1,k));axis image; colormap gray; 
title(['Filter ' num2str(k)])
end
print('OutputImages/Weights','-depsc');


%% Number clasify net Manual Spot Test
net.layers{end}.type = 'softmax';
imdbNum = load('matconvnet-1.0-beta25/data/mnist-baseline-simplenn/imdb.mat');
% The labels in the provided data set are wrong, we classify 0-9, not 1-10
% (ten is not a single digit number, the curator of this dataset needs to
% check their work)
net.meta.classes.name = arrayfun(@(x)sprintf('%d',x),0:9,'UniformOutput',false) ;

for ii = [1,2,3] %image numbers to be tested. 
    testim = imdbNum.images.data(:,:,1,ii);
    testim = single(testim);
    %testim = imresize(testim, net.meta.inputSize(1:2));
    
    %run it
    res = vl_simplenn(net,testim);
    
    %Classify Result
    scores = squeeze(gather(res(end).x));
    [bestScore, best] = max(scores);
    class= net.meta.classes.name{best};
    numberName=class;
    number = num2str(ii);
    figure(4); clf; imagesc(testim+imdbNum.images.data_mean); axis image; colormap gray; drawnow;
    title(sprintf('the number is %s, score is %.1f%%',numberName,bestScore*100))
    print(['OutputImages/MINST' number],'-depsc');
    
    
end


%% Even odd net Manual Spot Test
netEO.layers{end}.type = 'softmax';
imdbEO = load('matconvnet-1.0-beta25/data/mnist-evenodd-simplenn/imdb.mat');
% Classes didnt save correctly, make sure they are correct here
netEO.meta.classes.name={'Even', 'Odd'};
    
for ii = [1,6,8]
    testim = imdbEO.images.data(:,:,1,ii);
    testim = single(testim);
    %testim = imresize(testim, net.meta.inputSize(1:2));
    %run it
    res = vl_simplenn(netEO,testim);
    
    %Classify Result
    scores = squeeze(gather(res(end).x));
    [bestScore, best] = max(scores); % This is the correct index into the class names cell array on the line below (by itself this number is confusing)
    class= netEO.meta.classes.name{best}; % now index
    numberName=class;
    number = num2str(ii);
    figure(3); clf; imagesc(testim+imdbEO.images.data_mean); axis image; colormap gray; drawnow;
    title(sprintf('the number is %s, score is %.1f%%',numberName,bestScore*100))
    print(['OutputImages/EvenOdd' number],'-depsc');
    
end

%% Preprocessing images to be fed into the system

testImage = imread('InputImages/HRTest.png');
if size(testImage,3) == 3
    testImage = rgb2gray(testImage);
end
smallestSize = 20;
testImage = testImage/max(max(testImage));
imageThresholded = testImage < 0.01;
labelImage = ones(size(testImage,1),size(testImage,2));
[CCImage, endingLabelValue] = connectedComponentBZ(imageThresholded,1,labelImage);
CCImage= topologicalDenoising(endingLabelValue, CCImage, smallestSize);
targetSize = [28,28];
for ii=1:endingLabelValue
    threshold_image = CCImage ==ii;
    [x,y]=find(CCImage==ii);
    if ~(isempty(x)) 
        EDGE_BIAS = 65;
        Lboundx = min(x)-EDGE_BIAS;
        if Lboundx <=0
            Lboundx =1;
        end
        Lboundy = min(y)-EDGE_BIAS;
        if Lboundy <=0
            Lboundy =1;
        end
        Hboundx = max(x)+EDGE_BIAS;
        if Hboundx > size(CCImage,1)
            Hboundx = size(CCImage,1);
        end
        Hboundy = max(y)+EDGE_BIAS;
        if Hboundy > size(CCImage,2)
            Hboundy = size(CCImage,2);
        end
        temp_var = strcat( 'image_',num2str(ii));
        temp_image = double(threshold_image(Lboundx:Hboundx,Lboundy:Hboundy));
        sourceSize = size(temp_image);
        [X_samples,Y_samples] = meshgrid(linspace(1,sourceSize(2),targetSize(2)), linspace(1,sourceSize(1),targetSize(1)));
        temp_image = interp2(temp_image, X_samples, Y_samples);
        temp_image = temp_image > 0.75;
        se = offsetstrel('ball',3,3);
        temp_image = conv2(temp_image,ones(3),'same') > 1;
        temp_image = 255*single(temp_image);
        temp_image = conv2(temp_image, [1,3,1;3,9,3;1,3,1]/25, 'same');
        temp_image = temp_image -imdbNum.images.data_mean;
        res = vl_simplenn(net,temp_image);
        
        %Classify Result
        scores = squeeze(gather(res(end).x));
        [bestScore, best] = max(scores);
        number = num2str(ii);
        figure(ii); clf; imagesc(temp_image); axis image; colormap gray; drawnow;
        title(sprintf('the number is %s, score is %.1f%%',net.meta.classes.name{best+1}-1,bestScore*100))
        print(['OutputImages/HandWrittenImage' number],'-depsc');
        temp_var_im = 'temp_image';
        images{ii} = temp_image;
        
    end
end



%% CIFAR dataset
figure()
cifar = load('matconvnet-1.0-beta25/data/cifar-lenet/imdb.mat');
catdog = struct();
catdog.images.data = cifar.images.data(:,:,:,cifar.images.labels == 4 | cifar.images.labels == 6);
catdog.images.labels = 1 + (cifar.images.labels(cifar.images.labels == 4 | cifar.images.labels == 6) == 6);
catdog.images.set = ones(size(catdog.images.labels)); % Just use everything for training.  Not the best choice in the world
catdog.meta.classes={'cat','dog'};

for ii = [1,2,3]
    netC.layers{end}.type = 'softmax';
    testim = catdog.images.data(:,:,:,ii);
    testim = single(testim);
    testim = reshape(testim, netC.meta.inputSize);
    
    %run it
    res = vl_simplenn(netC,testim);
    
    %Classify Result
    scores = squeeze(gather(res(end).x));
    [bestScore, best] = max(scores);
    class= catdog.meta.classes{best};
    animalName=class;
    animal = num2str(ii);
    figure(ii); clf; imagesc(testim); axis image; colormap gray; drawnow;
    title(sprintf('the animal is %s, score is %.1f%%',animalName,bestScore*100))
    print(['OutputImages/CIFAR' animal],'-depsc');
end