What is CNN in Deep Learning? Architecture, Working & Applications

 Convulutional Neural Network(CNN)

CNN is Extension of MLP(Multi Layer Perceptron)

What is CNN?

CNN is a specialize neural network which mainly designed for image processing and visula pattern recognation and computer vision work.

Why CNN is important?

• Handle large image datasets efficiently
• Automatically extracts features
• Reduces number of parameters
• Provides high accuracy in vision tasks

Structure of CNN:

• Input layer: Receive image
• Convolution layer: 
• Feature detection using filter/cornel
• Detect edge, corner, texture
• Scane full image using filter slide
• Active function (ReLU):
• ReLU(Rectified Linear  Unit) f(x) = max(0, x)
• Add non-linearity
• Make zero of negative value
• Polling layer:
• Max polling / average pooling 
• Dimension reduction
• Reduce competition
• Prevent overfitting
• Fully connected layer
• Like traditional neural network
• Make final classification
• Softmax layer
• Used in  classification tasks

Real world applications of CNN

• Face recognition (Facebook, iPhone Face ID)
• Object detection (YOLO, R-CNN)
• Medical imaging (tumor detection)
• Self-driving cars
• Security surveillance
• Handwriting recognition
• License plate detection
• Agriculture disease detection
• Retail product recognition
• Satellite image analysis

Popular CNNs

1. LeNet-5 (1998): 

Proposed by: Yann LeCun

• One of the earliest CNN architectures
• Designed for handwritten digit recognition (MNIST)
• Uses convolution + pooling + fully connected layers
• Simple but historically important

Use Case: Bank cheque digit recognition

2. AlexNet (2012)

Developed by: Alex Krizhevsky

• Winner of ImageNet 2012 competition
• Much deeper than LeNet
• Introduced ReLU activation
• Used GPU training

Impact: Started the deep learning revolution in computer vision.

3.  VGGNet (2014)

Developed by: Oxford Visual Geometry Group

• Variants: VGG16, VGG19
• Uses very small 3×3 filters
• Deep but simple architecture

Pros: High accuracy

Cons: Very heavy (large parameters)

4. GoogLeNet / Inception (2014)

Developed by: Google

• Introduced Inception module
• Performs multiple convolutions in parallel
• More efficient than VGG

Advantage: High performance with fewer parameters

5. ResNet (2015)

Developed by: Microsoft

• Introduced Residual Connections (Skip Connections)
• Solved vanishing gradient problem
• Can train very deep networks (50, 101, 152 layers)

Famous Variants:

• ResNet50
• ResNet101

6. DenseNet (2017)

• Each layer connects to every other layer
• Improves feature reuse
• Reduces vanishing gradient

Benefit: Efficient parameter usage

7. MobileNet (2017)

Designed for: Mobile & embedded devices

• Lightweight CNN
• Uses depthwise separable convolution
• Fast and low computation

Use Case: Mobile vision apps

8. EfficientNet (2019)

• Scales depth, width, and resolution systematically
• Achieves high accuracy with fewer parameters

Popular Variants:

EfficientNet-B0 → B7

9. YOLO (You Only Look Once)

• Real-time object detection CNN
• Processes image in one pass
• Very fast