Convolutional
Neural
Network
- Sunny Chandra
Why CNN?
It helps to enable machines or computers
to have a vision system close to humans.
It gives them an ability to classify, identify objects and take decisions accordingly.
Let's observe our story of vision
We have been training our vision system
(i.e. neural networks associated with it)
since we were born.
Birth
Now
t
Demo
General perspective
Artisitic perspective
Can we achieve human like vision system with computers using plane programming?
One solution can be
use ANN
But not recommended for larger images
What's the solution then?
CNN
How ??
Convolution in terms of signals
In a layman's term, It's kind of mixing of 2 signals to produce a third signal.
This derived signal reveals some new info about the original signal itself.
Convolution in terms of images
When we convolve one image with
a filter (2D arrays) it results in a new image.
This new image may reveal about the
-
object boundaries
-
change in illumination, etc
present in the image
image
(6,6,3)
after ReLU
(4,4,5)
after
Pooling
(2,2,5)
Convolution layer
ReLU
Max Pooling
image
maps
(4,4,5)
Summary of Conv Layer
- P = amount of zero padding
- S = Stride
- F = Spatial extent
- K = No of Filters
Summary of Pooling Layer
- S = Stride
- F = Spatial extent
References
CNN architecture
By Sunny
