Lecture 6 – Convolutional Neural Networks (CNNs)

Introduction

Convolutional Neural Networks (CNNs) are the foundation of modern computer vision and one of the most important architectures in deep learning. CNNs excel at understanding images by automatically detecting edges, textures, shapes, and full objects, making them ideal for everything from face recognition to MRI analysis.

Traditional fully-connected networks fail with high-dimensional images, but CNNs solve this through convolutions, parameter sharing, and local receptive fields, making them efficient and powerful.

What Are Convolutional Neural Networks?

A CNN is a deep learning model designed to process grid-structured data, especially images.
Instead of manually defining features, CNNs automatically learn hierarchies:

• Layer 1: edges
• Layer 2: patterns/textures
• Layer 3: object parts
• Layer 4: complete objects

This makes CNNs far superior to classical machine learning in image-based tasks.

CNN Architecture (Core Building Blocks)

a) Convolution Layer

This is the heart of CNNs.
A convolution layer uses kernels (filters) that slide across the image and extract features.

Learns patterns like:

• vertical edges
• horizontal edges
• corners
• textures
• color gradients

Convolution Operation:

FeatureMap=Image∗Kernel

Each filter generates a new feature map.

Activation Function (ReLU)

After convolution, CNNs apply ReLU to introduce non-linearity so the network can learn complex patterns.

ReLU Example:

ReLU(x)=max⁡(0,x)

Lecture 5 – Optimization Algorithms in Deep Learning: SGD, Momentum, RMSProp & Adam Explained

Pooling Layer (Downsampling)

Pooling reduces spatial size & prevents overfitting.

Types:

• Max Pooling → selects the highest value
• Average Pooling → takes average

Reduces computation
Extracts dominant features
Improves translation invariance

Fully Connected Layer

After feature extraction, the image is flattened and fed into a classic neural network for classification.

Example:

• Cat probability = 0.93
• Dog probability = 0.07

https://www.tensorflow.org/tutorials

Why CNNs Work So Well?

• Local receptive fields → learn local patterns
• Weight sharing → fewer parameters
• Translation invariance → object recognized anywhere
• Scalability → works from small images to 4K data
• Automatic feature engineering

CNNs removed the need for manual feature design.

Real-World Applications of CNNs

Domain	Application
Healthcare	Tumor detection, CT/MRI analysis
Security	Face recognition (FaceID), CCTV surveillance
Automotive	Self-driving car vision systems
Social Media	Filters, image enhancement
Retail	Barcode/QR detection
Satellite	Land classification, weather imaging

Step-by-Step Example: Dog vs Cat Classifier

Step 1 – Input Image

Image resized to 128×128 pixels.

Step 2 – Convolution Layer

Extracts edges (whiskers, ears).

Step 3 – Pooling

Compresses size → keeps important features.

Step 4 – More Convolution Layers

Detects higher-level patterns (eyes, nose shape).

Step 5 – Fully Connected Layer

Combines all features → predicts category.

Step 6 – Softmax Output

Dog: 0.91
Cat: 0.09

Summary

Convolutional Neural Networks automatically learn visual features through convolution, pooling, and dense layers. They reduce computation, extract meaningful patterns, and achieve exceptional accuracy in modern image-based AI tasks.

People also ask: