What You Will Learn

In this article, you will learn:

• Why normal neural networks are not ideal for images.
• What a Convolutional Neural Network is.
• How filters, feature maps, pooling, and classification layers work.
• Where CNNs are used in real applications.
• Common interview questions about CNNs.

Introduction

Images are not simple rows of numbers. They contain spatial patterns such as edges, corners, shapes, textures, and objects.

A CNN, or Convolutional Neural Network, is a deep learning architecture designed to learn these spatial patterns.

CNNs are widely used for:

• Image classification.
• Object detection.
• Face recognition.
• Medical image analysis.
• Document scanning.
• Defect detection in manufacturing.

Why Images Need CNNs

A color image is usually represented as:

Height x Width x Channels

For example:

224 x 224 x 3

The 3 channels usually represent red, green, and blue.

If we send every pixel directly into a normal dense neural network, the model loses important spatial relationships and becomes very large. CNNs solve this by looking at small regions of the image at a time.

CNN Architecture

flowchart LR
    A["Input image"] --> B["Convolution"]
    B --> C["Activation"]
    C --> D["Pooling"]
    D --> E["More convolution blocks"]
    E --> F["Flatten"]
    F --> G["Dense layers"]
    G --> H["Prediction"]

Convolution Layer

A convolution layer applies small filters to the image.

A filter is a small matrix, often:

3 x 3
5 x 5

The filter slides over the image and detects patterns such as:

• Vertical edges.
• Horizontal edges.
• Curves.
• Corners.
• Textures.

Feature Maps

When a filter detects a pattern, it produces a feature map.

Image + Filter = Feature Map

Early CNN layers detect simple patterns. Deeper layers detect more complex patterns.

Layer Depth	Learns
Early layers	Edges and colors
Middle layers	Shapes and textures
Deep layers	Objects and object parts

Activation Function

CNNs commonly use ReLU after convolution.

ReLU(x) = max(0, x)

ReLU helps the network learn non-linear patterns and removes negative values that are not useful for activation.

Pooling Layer

Pooling reduces the size of feature maps while keeping important information.

The most common pooling type is max pooling.

Take the strongest value from each small region

Benefits:

• Reduces computation.
• Helps the model focus on strong features.
• Makes the model less sensitive to small shifts in the image.

Flatten and Dense Layers

After convolution and pooling, the model flattens the learned features into a vector.

Then dense layers use that vector to make a final prediction.

Example:

Image -> CNN features -> Dense layer -> Cat, dog, car, document

Real-World Example

For an insurance claim system, a CNN can classify uploaded vehicle images:

Input: Car damage photo
Output: Front bumper damage

This helps automate claim triage and route cases to the right team.

Common CNN Terms

Term	Meaning
Filter or kernel	Small matrix used to detect patterns
Stride	How far the filter moves each step
Padding	Extra border added around an image
Feature map	Output created by a filter
Pooling	Downsampling operation
Flattening	Converting feature maps into a vector

Interview Questions

What is a CNN?

A CNN is a neural network designed for image and spatial data. It uses convolution filters to learn patterns such as edges, shapes, and objects.

Why is CNN better than a dense network for images?

CNNs preserve spatial relationships, reuse filters, reduce parameters, and learn local patterns efficiently.

What is pooling?

Pooling reduces feature map size and keeps the strongest or average features from small regions.

Summary

CNNs are the foundation of modern computer vision. They learn visual patterns through convolution, activation, pooling, and classification layers.

Next, learn how sequence models such as RNNs and LSTMs process ordered data like text, time series, and speech.