Introduction to CNN and Its Applications in Computer Vision

 

Introduction to CNN and Its Applications in Computer Vision

1. Introduction

Humans can easily recognize objects, faces, animals, vehicles, handwritten text, and different activities in images. For example, when we see an image of a cat, we can identify it even when its position, size, colour, background, or direction changes.

For a computer, an image is not directly understood as a cat, person, car, or building. A computer reads an image as a collection of numerical pixel values.

A Convolutional Neural Network, commonly called a CNN, is a deep learning model specially designed to process and understand visual data such as images and videos.  

CNN is a type of deep neural network that automatically identifies important visual features and uses them for image classification, detection, recognition, and other computer vision tasks. 

What Is Computer Vision?

Computer Vision is a branch of Artificial Intelligence that enables computers to extract meaningful information from images and videos.

The main objective of computer vision is to help machines perform visual tasks that humans perform using their eyes and brain.

Examples

• Recognizing a person's face
• Detecting a vehicle in traffic
• Identifying a disease from a medical image
• Reading handwritten digits
• Detecting defective products
• Identifying objects in CCTV footage

Computer Vision enables a computer to see, analyse, and understand visual information.  

3. How Does a Computer Read an Image?

A digital image consists of small elements called pixels.

Each pixel contains a numerical value representing colour or brightness.

Grayscale Image

A grayscale image contains pixel values between:

$$0$$

Where:

• 0 represents black
• 255 represents white
• Values between 0 and 255 represent different shades of grey

For example:

This matrix represents the pixel values of a small grayscale image.

Colour Image

A colour image generally contains three channels:

• Red
• Green
• Blue

This is called an RGB image.

Each pixel contains three values:

$$(R,G,B)$$

Therefore, a computer processes images as numerical matrices rather than directly understanding the objects present in them.

---

4. What Is a Convolutional Neural Network?

A Convolutional Neural Network is a specialised deep learning architecture used mainly for processing image data.

A CNN automatically learns visual features such as:

• Edges
• Lines
• Corners
• Curves
• Textures
• Shapes
• Object parts
• Complete objects

A traditional machine learning model may require manual feature extraction. In CNN, feature extraction is performed automatically during training.

Basic Working

Input Image

     ↓

Feature Extraction

     ↓

Important Features

     ↓

Classification

     ↓

Predicted Object   

For example:

Input Image → Edges → Shapes → Cat Features → Cat Prediction

 

5. Why Is CNN Used for Images?

Images usually contain a very large number of pixel values.

Consider a colour image of size:

$$224\times 224\times 3$$

The total number of input values is:

$$= 150,528$$

If all these pixels are directly connected to every neuron in a traditional Artificial Neural Network, the number of parameters becomes extremely large.

This creates several problems:

• High computational cost
• Large memory requirement
• Longer training time
• Increased possibility of overfitting
• Loss of spatial information

CNN solves these problems by examining small regions of an image and sharing the same feature detector across different image locations.

Difference Between ANN and CNN

Artificial Neural Network	Convolutional Neural Network
Used mainly for structured or tabular data	Used mainly for images and visual data
Input is generally a one-dimensional feature vector	Input can be a two-dimensional or three-dimensional image
Every neuron may connect with all input values	Neurons examine small local regions
Requires a large number of parameters for images	Uses fewer parameters through weight sharing
Does not naturally preserve spatial relationships	Preserves spatial relationships between pixels
Feature extraction may be performed manually	Features are learned automatically
Suitable for prediction and classification	Suitable for image recognition, detection and segmentation

7. What Does Spatial Information Mean?

Spatial information describes the position and relationship of pixels within an image.

For example, in a face:

• Eyes are generally above the nose.
• The nose is generally above the mouth.
• The eyes are located near each other.
• The mouth is located in the lower part of the face.

The position of these features is important for recognizing the object.

A CNN preserves these relationships while processing an image.

Spatial information tells the model where visual features are located and how they are related to one another.

  

8. How Does CNN Recognize an Object?

A CNN learns image features gradually through multiple layers.

Consider the classification of a human face.

Initial Layers

The first few layers may identify simple features such as:

• Horizontal edges
• Vertical edges
• Curved edges
• Brightness changes

Middle Layers

The middle layers may combine simple features to identify:

• Eyes
• Nose
• Mouth
• Ears

Deeper Layers

The deeper layers combine all these features to recognize:

• A complete face
• A particular person
• A facial expression

The process can be represented as: 

Pixels

   ↓

Edges

   ↓

Lines and curves

   ↓

Object parts

   ↓

Complete object

   ↓

Prediction 

This is known as hierarchical feature learning.

---

9. Hierarchical Feature Learning

CNN does not identify a complete object in a single step. It learns features at different levels.

CNN Layer Level	Features Learned
Initial layers	Edges, lines and colour patterns
Middle layers	Curves, textures and simple shapes
Deeper layers	Eyes, wheels, ears and object components
Final layers	Complete objects such as a person, car, dog or tumour

Simple Example

For a car image:

Pixels → Edges → Curves → Wheels and windows → Complete car

For a face image:

Pixels → Lines → Eyes and nose → Facial structure → Face  

10. Why Is It Called a Convolutional Neural Network?

It is called a Convolutional Neural Network because it uses a mathematical operation called convolution.

During convolution, a small matrix called a filter or kernel moves across the image and identifies important features.

For example, a filter may identify:

• Vertical edges
• Horizontal edges
• Curves
• Textures
• Colour patterns

The detailed convolution operation will be studied in the next section. 

11. Basic CNN Architecture

A CNN generally contains the following layers:

Input Image
      ↓
Convolution Layer
      ↓
Activation Function
      ↓
Pooling Layer
      ↓
Flatten Layer
      ↓
Fully Connected Layer
      ↓
Output Layer 

Role of Each Layer

Layer	Main Function
Input Layer	Receives the image
Convolution Layer	Extracts important image features
Activation Layer	Introduces non-linearity
Pooling Layer	Reduces feature-map size
Flatten Layer	Converts feature maps into a one-dimensional vector
Fully Connected Layer	Performs classification
Output Layer	Produces the final prediction

Applications of CNN in Computer Vision

12. Image Classification

Image classification assigns one class label to an entire image.

Examples

• Cat or dog
• Benign or malignant tumour
• Healthy or diseased plant
• Car, bus or motorcycle
• Normal or defective product

Input Image → CNN → Predicted Class 

For example:

Chest X-ray → CNN → Normal or Pneumonia

---

13. Object Detection

Object detection identifies:

1. Which objects are present
2. Where the objects are located

It usually draws a rectangular box called a bounding box around each detected object.

Examples

• Detecting people in CCTV footage
• Detecting vehicles on roads
• Detecting helmets worn by riders
• Detecting tumours in medical images
• Detecting products on store shelves  

Difference from Classification

Image Classification	Object Detection
Predicts one label for an image	Identifies multiple objects
Does not show object location	Shows object location using bounding boxes
Example: “This image contains a car”	Example: “There are three cars at these locations”

---

14. Image Segmentation

Image segmentation divides an image into meaningful regions.

Instead of drawing only a rectangular box, segmentation identifies the exact pixels belonging to an object.

Examples

• Identifying the exact tumour region
• Separating roads, vehicles and pedestrians
• Identifying damaged regions in machinery
• Separating land, water and vegetation in satellite images
• Identifying organs in medical images

Input Image → CNN → Pixel-level object regions

---

15. Face Recognition

CNNs are widely used for recognising and verifying human faces.

Applications

• Smartphone face unlocking
• Attendance systems
• Airport security
• Employee verification
• Surveillance systems
• Banking authentication

CNN extracts facial features and compares them with stored facial information.

---

16. Medical Image Analysis

CNNs can analyse medical images such as:

• X-rays
• Computed Tomography scans
• Magnetic Resonance Imaging scans
• Ultrasound images
• Mammograms
• Skin lesion images

Applications

• Breast cancer detection
• Brain tumour detection
• Pneumonia detection
• Lung disease classification
• Skin cancer detection
• Diabetic retinopathy detection

CNN assists doctors by identifying patterns that may indicate disease. It supports medical decision-making but does not independently replace medical professionals.

---

17. Autonomous Vehicles

Self-driving vehicles use CNNs to understand road scenes.

CNNs can identify:

• Roads
• Traffic signals
• Pedestrians
• Vehicles
• Road signs
• Lane markings
• Obstacles

Example:

Camera Image → CNN → Road objects → Driving decision

---

18. Handwriting Recognition

CNNs are used to recognize handwritten characters and digits.

Applications

• Reading postal codes
• Processing bank cheques
• Digitizing handwritten forms
• Recognizing handwritten numbers
• Optical Character Recognition systems

The MNIST handwritten-digit dataset is commonly used to introduce CNN-based image classification.

---

19. Industrial Quality Inspection

CNNs can automatically inspect products in manufacturing industries.

Applications

• Detecting cracks
• Identifying scratches
• Finding missing parts
• Identifying incorrect shapes
• Detecting surface defects
• Separating defective products

This improves production quality and reduces dependence on manual inspection.

---

20. Agriculture

CNNs are used in smart agriculture for:

• Plant disease detection
• Fruit classification
• Crop monitoring
• Weed identification
• Pest detection
• Fruit quality assessment

For example:

Leaf Image → CNN → Healthy or Diseased

---

21. Security and Surveillance

CNNs help analyse images and videos collected through surveillance cameras.

Applications

• Person detection
• Suspicious activity detection
• Intrusion detection
• Crowd monitoring
• Weapon detection
• Vehicle-number recognition

---

22. Satellite and Remote-Sensing Images

CNNs can analyse satellite and aerial images.

Applications

• Land-use classification
• Forest monitoring
• Flood detection
• Fire detection
• Urban-development analysis
• Crop-area identification
• Disaster assessment

---

23. Real-Life Example

Consider a hospital where a mammogram image is submitted for breast cancer analysis.

The CNN processes the image through several layers:

Mammogram Image
       ↓
Edges and boundaries
       ↓
Textures and shapes
       ↓
Suspicious tumour patterns
       ↓
Benign or Malignant prediction

The initial layers detect basic edges and textures. The deeper layers identify complex tumour-related patterns. Finally, the fully connected and output layers generate the classification result.

---

24. Advantages of CNN

• Automatically extracts relevant features
• Preserves spatial relationships in images
• Requires less manual feature engineering
• Uses fewer parameters than a fully connected network for image processing
• Provides high performance in visual-recognition tasks
• Can learn simple as well as complex image patterns
• Can be applied to images, videos and medical scans
• Supports classification, detection and segmentation

---

25. Limitations of CNN

• Requires a sufficiently large dataset
• Training can require high computational power
• Complex CNNs may require a Graphics Processing Unit
• Model training may take considerable time
• Performance depends on the quality of the images
• CNN decisions may be difficult to interpret
• Incorrect or biased training data can produce unreliable predictions
• Hyperparameters and architecture must be carefully selected