Anomaly Detection: Identifying Rare and Unusual Patterns in Data

👁️ Optical Character Recognition (OCR)

Sliding Windows and Photo OCR

Photo OCR stands for:

Photo Optical Character Recognition

Goal:

Image → Detect Text → Read Characters → Final Text

Applications:

• Google Lens
• document scanning
• self-driving cars reading signs
• helping visually impaired users
• searching photos by text

---

Sliding Window Detection 🖼️

A small rectangle moves across the image.

flowchart LR
    A[Image Patch 1] --> B[Classifier]
    C[Image Patch 2] --> B
    D[Image Patch 3] --> B
    B --> E[Pedestrian Yes or No]

At each location:

1. crop image patch
2. resize if needed
3. run classifier
4. move window

We train a classifier

$$Image\ Patch \rightarrow y \in \{0,1\}$$

Where

1. Positive Examples ($1$)

Contains pedestrian

2. Negative Examples ($0$)

No pedestrian

Stride / Step Size

The amount the window moves each time.

Example:

• stride = 1 pixel
• stride = 4 pixels
• stride = 8 pixels

Small stride

• More accurate detection
• Slower

Large stride

• Faster detection
• Low accuracy: May miss objects

Multi-Scale Detection

Objects can appear at different sizes.

So we use:

• small windows
• medium windows
• large windows

flowchart TD
    A[Small Window]
    B[Medium Window]
    C[Large Window]

Each patch is resized before classification.

Your input images are 1000x1000 pixels.

• Sliding windows detector sizes: 10x10 , 20x20
• Stride 2

Number of iteration per side= 1000/2 = 500

Total iteration = 500X 500 = 250000

For 2 scales = 2 X 250000 = 500,000 times

---

Machine Learning Pipeline

Photo OCR is built as a pipeline of smaller ML systems.

Benefits:

• easier debugging
• modular design
• easier teamwork

Each stage solves one smaller problem.

flowchart TD
    A[Input Image] --> B[Text Detection]
    B --> C[Character Segmentation]
    C --> D[Character Recognition]
    D --> E[Final Text Output]

1. Text Detection

Find where text exists inside the image.

Now apply sliding windows to text.

Train classifier on:

1. Positive Examples

Image patches containing text

2. Negative Examples

Image patches without text

Text Detection Process

flowchart TD
    A[Input Image]
    --> B[Slide Window Across Image]
    --> C[Classifier Predicts Text Probability]
    --> D[Probability Heatmap]
    --> E[Bounding Boxes Around Text]

Classifier outputs:

White  → high confidence text
Gray   → uncertain
Black  → no text

Expansion Operator

After detection, nearby white regions are expanded.

Purpose:

• merge neighboring text pixels
• form larger connected text regions

flowchart LR
    A[Small White Blobs]
    --> B[Expansion]
    --> C[Larger Connected Regions]

Connected Components

Now detect connected white regions and draw boxes.

┌─────────────┐
│ Antique Mall│
└─────────────┘

We also filter weird shapes.

Aspect Ratio Filtering

Text regions are usually:

Wide > Tall

So discard:

• tall thin regions
• random noisy blobs

---

2. Character Segmentation

Split the detected text into individual characters.

Example:

ANTIQUE

Desired output:

A | N | T | I | Q | U | E

1D Sliding Window

Now the window moves only horizontally.

flowchart LR
    A[Character Strip]
    --> B[Slide Left to Right]
    --> C[Predict Split Locations]

Classifier predicts:

Should we split here?

Training Character Split Classifier

1. Positive Example

A | N

Correct split location.

2. Negative Example

AN

Inside one character.

---

3. Character Recognition

Classify each character. Image of character → Predicted Letter

• [A-image] → "A"
• [N-image] → "N"

Now each segmented character becomes a classification problem.

flowchart LR
    A[Character Image]
    --> B[Classifier]
    --> C[A-Z or 0-9]

Example:

Image of "A" → Predict "A"

This is a multiclass classification problem.

Artificial Data Synthesis

Generate fake training data to creates huge datasets cheaply.

Example:

• different fonts
• rotations
• shadows
• blur

---

Full OCR System

flowchart TD
    A[Photo]
    --> B[Sliding Window Text Detection]
    --> C[Text Bounding Boxes]
    --> D[Character Segmentation]
    --> E[Character Classification]
    --> F[Combine Characters]
    --> G[Final Readable Text]

---

Modern Computer Vision

Classical sliding windows are historically important.

Modern systems now often use:

• CNNs
• YOLO
• Faster R-CNN
• SSD
• Vision Transformers

These are much faster and more accurate.

---

Core Takeaway

Large machine learning systems are usually:

Many small ML models
working together
inside a pipeline