Computer Vision &

Its Applications

By
Dr.Eng.Wafaa Shalash
Objectives
• Basics of CNN.
• Different CNN Layers structure
• Using deep learning (hands-on
examples)
• Investigating topics related to deep
learning:
• CNN Architecture
• preparing data
• data augmentation
• building CNN from scratch
• optimizers
Deep Learning
To Build a Deep Model We Need:

Deep Net
Data Design

Training
Neural Networks Vs. Deep Learning
Deep Learning
Deep Network Output Example

F F
u u
l l
l l
y y
C C S
o o o C
n n f l
n n t
e e a
M s
c c
a s
t t
e e x
d d
 How does
Convolution Neural
Networks work?
Step 1: Convolute input with mask
How does
Convolution Neural • Step 1: Convolute input with mask
Networks work?
Convolution with
Padding Padding
Convolution with
Padding Padding
Convolution with
Striding
Example: stride by 2
Convolution with
Striding
Example: stride by 2
Convolution with RGB Channels
Input
Convolution with RGB Channels Input

Each of the per-channel processed versions are then summed together to

form one channel.
Step 2 : The Rectified Linear
Unit (RELU)

activation function
A function (for example, ReLU or sigmoid)
that takes in the weighted sum of all of the
inputs from the previous layer and then
generates and passes an output value
(typically nonlinear) to the next layer.
 (Sigmoid)

How to Choose an Activation

Function for Deep Learning
• Rectified Linear Activation (ReLU)
• Logistic (Sigmoid)
• Hyperbolic Tangent (Tanh) (ReLU)
• Multilayer Perceptron (MLP): ReLU activation
function.
• Convolutional Neural Network (CNN): ReLU
activation function.
(Tanh)
• Recurrent Neural Network: Tanh and/or Sigmoid
activation function.
Step 3 : Pooling
Many rules could eb applied in
pooling
Max. pooling
Min. Pooling
Average Pooling
 Step 3 : Pooling
Many rules could eb applied
in pooling
Max. pooling
Min. Pooling
Average Pooling
A Convolutional Module
Fully
Connected
Layer
Example of an end-to-end
structure of a convolutional
neural network.
Fully CNN Example 1.
Example of an end-to-end
structure of a
convolutional neural Fully CNN Example 2.
network.
 • Optimizer type
• Learning Rate
Training • Batch size
• GPU usage

Net
Design
• Select Layers
• Select Activation
Fn.
Data

• Labelling
• Augmented
• Splitting
 Approaches to
Design a CNN

Reuse a pretrained
Design Layers from
Network with
Scratch.
Transfer Learning
 Design Layers from Scratch.
F F
u u
ll ll
y y
C C
S
o o
o
n n C
f
n n l
t
e e a
M
c c s
a
t t s
x
e e
d d

2D Convolutional Layer
ReuL
Pooling
Practical
Example with
MATLAB
 How to design
a CNN with
MATLAB

Use MATLAB
Write the
Deep Network
code directly
Designer
 30

Coding From
Scratch
Preparing Data
1. Labeling
2. Splitting
3. Augmentation

31
 MATLAB Data
Store
Look at data and store information
about it, not loading it into
memory

32
 MATLAB Data
Store
Datastorte object manage memory and Look only the
Required data into memory

33
Image Browser
is a Tool from
MATLAB App
to Brows
Image Data

34
 You can
Brows
individual
image
properties

12/11/2022 35
Augmentation

36
 Splitting Data

Image
Data

Training Validation Testing

Data Data Data

12/11/2022 37
 MATLAB Code : Import Data

digitDatasetPath = 'flower_photos’; Data Location

imds = imageDatastore(digitDatasetPath, ...

'IncludeSubfolders',true,'LabelSource','folde
rnames');
Data Lable

38
 MATLAB Code: Split Data
[trainImgs,testImgs, ValidateImg] = splitEachLabel(imds,0.7,0.2,'randomized');

39
 MATLAB Code
• digitDatasetPath = 'flower_photos';

• imds = imageDatastore(digitDatasetPath, ...

'IncludeSubfolders',true,'LabelSource','foldernames');

• [trainImgs,testImgs, ValidateImg] =
splitEachLabel(imds,0.7,0.2,'randomized');

40
(MNIST
dataset)

41
Example 1: Load and
Explore Data

digitDatasetPath = Minest_dataset;
imds = imageDatastore(digitDatasetPath, ...
'IncludeSubfolders',true,'LabelSource','foldernames’);
figure;
perm = randperm(10000,20);
for i = 1:20
subplot(4,5,i);
imshow(imds.Files{perm(i)});
end
Define Network Architecture
labelCount = countEachLabel(imds)
img = readimage(imds,1); size(img)
numTrainFiles = 750;
[imdsTrain,imdsValidation] = splitEachLabel(imds,numTrainFiles,'randomize’);
layers = [ imageInputLayer([28 28 1])
convolution2dLayer(3,8,'Padding','same’)
batchNormalizationLayer
reluLayer
maxPooling2dLayer(2,'Stride',2)
convolution2dLayer(3,16,'Padding','same’)
batchNormalizationLayer
reluLayer
maxPooling2dLayer(2,'Stride',2)
convolution2dLayer(3,32,'Padding','same’)
batchNormalizationLayer
reluLayer
fullyConnectedLayer(10)
softmaxLayer
classificationLayer];
 Network
Analyzer

net2= AlexNet;
analyzeNetwork(net)

plot(layerGraph(layers))
 Training Options and Training

options = trainingOptions('sgdm', ...

'InitialLearnRate',0.01, ...
'MaxEpochs',4, ...
'Shuffle','every-epoch', ...
'ValidationData',imdsValidation, ...
'ValidationFrequency',30, ...
'Verbose',false, ...
'Plots','training-progress’);
net = trainNetwork(imdsTrain,layers,options);
 Testing and Confusion Matrix

testPreds = classify(net,imdsValidation)
acc = (nnz(testPreds ==
imdsValidation.Labels)/length(testPreds))*100;
confusionchart(imdsValidation.Labels,testPreds);
Hands- on with
MATLAB
Use MATLAB
Deep Network
Designer
Use MATLAB Deep Network Designer
Use MATLAB
Deep Network
Designer
Hands- on with
MATLAB
Adjust Training Options
Adjust Training Parameters
(Tunning)
Training Process is like the process to find the minimum among mountains.
The main target is to minimize the Loss function.
Learning Rate
• Small learning rate makes you arrive slowly.

• Large learning rate might make jumping

around the minimum and couldn’t reach it.
Optmizer or Solvers
Solver for training network, specified as one of the following:
•'sgdm' — Use the stochastic gradient descent with momentum
(SGDM) optimizer. You can specify the momentum value using
the 'Momentum' name-value pair argument.
•'rmsprop'— Use the RMSProp optimizer. You can specify the decay
rate of the squared gradient moving average using
the 'SquaredGradientDecayFactor' name-value pair argument.
•'adam'— Use the Adam optimizer. You can specify the decay rates of
the gradient and squared gradient moving averages using
the 'GradientDecayFactor' and 'SquaredGradientDecayFactor' name-value
pair arguments, respectively.
Comparison between different
Optimizers

Accuracy Sedm
Adam
Rmsprop

Iterations
Comparison between different Optimizers

Sedm
Loss Adam
Rmsprop

Iterations
Optmizers
 • At each iteration, a subset of the training images, known as a mini-batch, is used to update the
weights. Each iteration uses a different mini-batch. Once the whole training set has been used,
What is a Mini- that's known as an epoch.
• The maximum number of epochs (MaxEpochs) and the size of the mini-batches (MiniBatchSize)
Batches? are parameters you can set in the training algorithm options.
• Note that the loss and accuracy reported during training are for the mini-batch being used in the
current iteration.
• By default, the images are shuffled once prior to being divided into mini-batches. You can control
this behavior with the Shuffle option.
Using GPUs
• A GPU (graphics processing unit) can significantly speed up the
many computations required for deep learning. If the computer
doesn't have a supported GPU, the training can be performed
on the CPU, but it may take a long time. If you're going to get
serious about deep learning, you'll want to train your network on
a computer with a GPU that can handle the processing.
• If you have an appropriate GPU and Parallel Computing Toolbox
installed, the trainNetwork function will automatically perform the
training on the GPU -- no special coding required.
• If not, the training will be done on your computer's CPU instead.
This gives you the option to experiment a little before committing
to purchasing the needed hardware and software.
Augmentation …
more
• Resize
• displacement
• Rotate
• Shear
 options = trainingOptions('adam',
'GradientDecayFactor',0.6,"LearnRateSchedule
",'piecewise','ExecutionEnvironment','gpu',
...
'LearnRateDropFactor',0.5, ...
'LearnRateDropPeriod',5, ...
Training 'SquaredGradientDecayFactor',0.6, ...

Options 'InitialLearnRate',0.0001, ...

'MaxEpochs',100,...
Example 'L2Regularization',0.099,...
'Shuffle','every-epoch',...
'ValidationData',validAug,...
'ValidationFrequency',20, ...
'MiniBatchs',65,...
'Plots','training-progress');
Drawbacks and Limitations of
Using Deep Learning for
Computer Vision

• AI technology presents some

limitations that prevent it from
becoming a mainstream security
tool.

• Resources—organizations need lots

of resources including data,
memory, and computing power.

• Data sets—Deep learning needs a

huge datasets to be trained.

64
Next time: Using Transfer Learning

65