Scribd
Upload
54 views12 pages

Structure of Convolutional Neural Networks - Deep Learning

Uploaded by

nirajaadithya.dasireddi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
54 views12 pages

Structure of Convolutional Neural Networks - Deep Learning

Uploaded by

nirajaadithya.dasireddi
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 12

Before we build a CNN…

Before we build a CNN, the image data we have must be pre-processed to ensure
it is ready to be used by the neural network.

Pre-processing of image data for CNNs involves steps like:

1) Re-sizing images

⟶ The images of different sizes are converted to one uniform size (usually to 224 x 224).

2) Normalizing the pixel data of images

⟶ Since pixels can have varying values of density, we normalize them to one range, usually
between 0 and 1, or -1 and +1.
3) Data Augmentation :

⟶ Data augmentation is the process of generating more training examples


(images) by rotating, flipping, lighting changes, and applying similar techniques
on existing images.

⟶ The training data is thus increased from existing data, and the CNN will
work on variations of the same image and generalize better.

Pre-processing of images can also be done within the neural network by using
the tf.keras.preprocessing.image module,

with which we can define some layers which process the images before they
reach the convolutional layer.
Structure of Convolutional Neural Networks

The structure of a convolutional neural network usually consists of the


following layers :

1) Convolutional Layers : Apply kernels to extract features from the image


and generate feature maps.

2) Pooling Layers : Extract information from feature maps and reduce their
size.

3) Flatten Layer : To convert multi-dimensional feature maps into 1-D


vectors.

4) Fully Connected Layer : To process the flattened image data and carry
out the classification.

! Batch Normalization Layer : To normalize inputs between layers.

! Dropout Layer : Drop out a fraction of neurons from a layer.

Let us examine the role of each layer in detail :


1) Convolution Layer

⟶ This layer applies kernels to images, extracts features from them and
generates feature maps.

https://learnopencv.com/understanding-convolutional-neural-networks-cnn/

⟶ It includes a layer with activation functions for the convolutional layer.

⟶ A convolution layer combined with batch normalization, pooling, dropout


layers is called a convolutional block.

⟶ Depending on the task, multiple convolutional blocks are required and


placed.
2) Batch Normalization Layer :

⟶ It normalizes the inputs between layers, can be placed before or after


the activation function layer.
! It is optional, and used in very deep neural networks, as a regularization technique.

https://towardsdatascience.com/batch-norm-explained-visually-how-it-works-and-why-neural-networks-need-it-b18919692739
3) Pooling Layer :

⟶ Extract information from feature maps and shrinks their size.

⟶ This layer can apply different types of pooling methods, like average
pooling, max pooling etc.

https://towardsai.net/p/l/introduction-to-pooling-layers-in-cnn

! Although it is technically optional, pooling layers are essential in almost every


case.
4) Dropout Layer :

⟶ This layer drops out a fraction of neurons from the previous layer, reducing
the number of units, to make the network learn more robust features.

⟶ It is used in deep CNNs, often with complex architectures.

https://www.researchgate.net/figure/Example-of-dropout-layer-probability-of-50-appears-on-the-right_fig8_333411007
5) Flatten Layer :

⟶ This layer flattens the multi-dimensional data from convolutional blocks to


one-dimensional vectors, preparing it for fully connected layers.

⟶ This one-dimensional vector suitable as input for fully connected layers,


which typically work with flat data for classification or regression tasks.

https://www.superdatascience.com/blogs/convolutional-neural-networks-cnn-step-3-flattening
6) Fully Connected Layers :

⟶ The actual classification or prediction happens in these layers.

⟶ The normal fully-connected layers takes input from the Flatten layer and
carries out the classification/prediction task.

⟶ The output layer, has the appropriate function like Softmax, Sigmoid etc.
based on the task, which is usually multi-class or binary classification.

In a manner of understanding,

the fully connected layers can be interpreted as separate neural network for which the input is the
feature maps generated by the convolutional portion of the network.
We combine all the parts sequentially
to build a complete convolutional neural
network.

Multiple convolution blocks, which will


extract features, then a set of fully
connected (Dense) layers which will use
those features to make classification.

Some examples :

https://www.mdpi.com/2071-1050/15/23/16292

For classifying images of horses, zebras and dogs :

https://www.linkedin.com/pulse/what-convolutional-neural-network-cnn-deep-learning-nafiz-shahriar/

https://www.researchgate.net/figure/Functional-structure-of-the-proposed-convolutional-neural-network-architecture-Conv_fig2_350284279

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy