Training CNN using Matlab R2017b NN and CV toolbox

train_network.m

% Load training data.
imageDir = fullfile('training_data');
labelDir = fullfile('label_data');


% Create an image datastore for the images.
imds = imageDatastore(imageDir, 'IncludeSubfolders',true, ...
    'LabelSource','foldernames');

% Create a pixelLabelDatastore for the ground truth pixel labels.
classNames = ["good","bad"];
labelIDs   = [255 0];
pxds = pixelLabelDatastore(labelDir, classNames, labelIDs);

%%%%%% Visualize training images and ground truth pixel labels.
I = read(imds);
C = read(pxds);
figure
I = imresize(I,5);
L = imresize(uint8(C),5);
imshowpair(I,L,'montage')

% Create a semantic segmentation network. This network uses a simple 
% semantic segmentation network based on a downsampling and upsampling 
% design.
numFilters = 40;
filterSize = 15;
numClasses = numel(categories(imds.Labels));
layers = [
    imageInputLayer([64 64 1])
    convolution2dLayer(filterSize,numFilters,'Padding',0)
    maxPooling2dLayer(2,'Stride',2)
    convolution2dLayer(filterSize,numFilters/2,'Padding',0)
    maxPooling2dLayer(2,'Stride',2)
    reluLayer()
    transposedConv2dLayer(filterSize,numFilters,'Stride',2);
    convolution2dLayer(1,numClasses);
    softmaxLayer()
    pixelClassificationLayer()
]

% Setup training options.
opts = trainingOptions('sgdm', ...
    'InitialLearnRate', 1e-3, ...
    'MaxEpochs', 50, ...
    'MiniBatchSize', 100);

% Create a data source for training data.
trainingData = pixelLabelImageSource(imds,pxds);


% Train the network.
net = trainNetwork(trainingData,layers,opts);
mem_net = net
save mem_net

% Read and display a test image.
testImage = imread('fil028.jpg');
figure
imshow(testImage)

% Segment the test image and display the results.
%C = semanticseg(testImage,net);
%B = labeloverlay(testImage,C);
%figure
%imshow(B)

%%%%%% Improve the results if necessary
% The network failed to property segment the triangles and classified every
% pixel as "background". The training appeared to be going well with 
% training accuracies greater than 90%. However, the network only learned 
% to classify the background class. To understand why this happened, you 
% can count the occurrence of each pixel label across the dataset.

% The majority of pixel labels are for the background. The poor results are
% due to the class imbalance. Class imbalance biases the learning process 
% in favor of the dominant class. That's why every pixel is classified as 
% "background". To fix this, use class weighting to balance the classes. 
% There are several methods for computing class weights. One common method 
% is inverse frequency weighting where the class weights are the inverse of
% the class frequencies. This increases weight given to under-represented 
% classes.

%tbl = countEachLabel(trainingData)

%totalNumberOfPixels = sum(tbl.PixelCount);
%frequency = tbl.PixelCount / totalNumberOfPixels;
%classWeights = 1./frequency

%layers(end) = pixelClassificationLayer('ClassNames',tbl.Name,'ClassWeights',classWeights);

%net = trainNetwork(trainingData,layers,opts);

% Try to segment the test image again.

C = semanticseg(testImage,net);
B = labeloverlay(testImage,C);
figure
imshow(B)