← BackGrades 5-6 | Chapter 2: AI in Action

How AI Sees Images 👁️

Computer Vision Deep Dive

Session 5: Image Recognition

Pixels: Building Blocks 🧩

Every image on your screen is made of tiny dots called pixels. AI doesn't see images like you do - it sees numbers!

How Pixels Work:

  • Each pixel is a tiny square with a color
  • Red image = high red number (0-255)
  • Blue image = high blue number
  • A 100x100 photo = 10,000 pixels to analyze
  • AI processes pixel-by-pixel data

Feature Detection 🔍

AI doesn't immediately recognize "this is a cat". Instead, it detects features - lines, curves, shapes.

How Feature Detection Works:

  1. AI looks for edges (where color changes)
  2. Detects corners and curves
  3. Groups features into shapes
  4. Recognizes patterns
  5. Compares to billions of examples it learned from

Cat Recognition Example 🐱

How does AI recognize a cat?

Step by Step:

  • Detect ears: Two triangular shapes at the top
  • Detect eyes: Two oval shapes with pupils
  • Detect whiskers: Long lines coming from the face
  • Detect fur texture: Specific pattern of pixels
  • Combine features: "This looks like a cat!"

Training Data: The Problem ⚠️

AI is only as good as the training data it learned from. This creates major issues!

What If Training Data Is Biased?

  • If AI learned from 90% orange cats, it might not recognize black cats
  • If trained mostly on clear photos, it fails on blurry images
  • If trained on one breed (Siamese), it might not recognize Maine Coons

Real-World Applications 🌍

Computer Vision Uses:

  • Face Recognition: iPhone Face ID, Facebook photo tagging
  • Medical Imaging: Detecting tumors in X-rays
  • Self-Driving Cars: Recognizing pedestrians, traffic signs
  • Agriculture: Identifying plant diseases
  • Retail: Checkout without scanning (Amazon Go stores)

Bias in AI Vision 🎨

Studies have shown AI vision systems have significant bias problems.

Real Examples of Bias:

  • Google Photos tagged African Americans as "gorillas"
  • Some facial recognition systems have 99% accuracy for white men, 35% for dark-skinned women
  • Autonomous vehicles trained in US don't recognize European road signs

Why this matters: If a self-driving car can't recognize you because of your skin color, it could cause accidents!

Improving AI Vision 🚀

Solutions:

  • Diverse training data: Include many types of images
  • Audit results: Test on different groups
  • Transparency: Tell people AI is being used
  • Human review: Have humans double-check decisions
  • Fairness testing: Regular bias assessments

What We Learned 🎓

  • AI sees images as pixel data (numbers)
  • AI detects features (edges, shapes, patterns)
  • AI compares features to training data
  • Biased training data creates biased AI
  • Computer vision has many applications but also risks
  • Diversity in data and human oversight are critical

AI Vision Explained! 🎉

AI sees differently than humans - with both advantages and risks

Next Session: How AI Understands Language