Module 6: Advanced Computer Vision – See Like AI

Duration: Week 12-13
Difficulty: Advanced
Prerequisites: Module 4 completed

—

🎯 What You’ll Learn

Build AI that can identify objects, detect faces, and understand images like self-driving cars do!

—

🏗️ Part 1: ResNet – Going Deeper Without Breaking

The Problem

Deeper networks should be better, but they often perform WORSE!
Why? Gradients vanish (become too small to learn)

ResNet’s Solution: Skip Connections

Regular: Input → Layer1 → Layer2 → Layer3 → Output
ResNet:  Input → Layer1 → Layer2 → Layer3 → Output
                  ↓_____________↑ (skip connection!)

Analogy: Like having shortcuts in a maze

• Information can flow directly
• Easier to learn
• Can build 50-100 layer networks!

Project: Build ResNet-18

• 18 layers deep
• Dataset: CIFAR-10 or your own photos
• Goal: 80%+ accuracy
• Training: 1-2 hours

—

🎨 Part 2: Data Augmentation – Getting More from Less

The Problem

Need 1000s of images, but only have 100?

Solution: Transform existing images!

Transformations:
1. Flip: Mirror image horizontally
2. Rotate: Turn 10-15 degrees
3. Crop: Random sections
4. Color: Adjust brightness, contrast
5. Noise: Add small random changes

Result: 100 images → 1000s of variations!

Project: Augmentation Pipeline

• Use Albumentations library
• Create 10 variations of each image
• Compare: With vs without augmentation
• Result: 10-20% better accuracy!

—

🎯 Part 3: Object Detection – Finding Things in Images

Classification vs Detection

• Classification: “This is a cat” ✓
• Detection: “There’s a cat at position (100, 200)” ✓✓

How YOLO Works (You Only Look Once)

1. Divide image into grid (e.g., 7×7)
2. Each cell predicts:
– What object? (cat, dog, car)
– Where is it? (x, y, width, height)
– How confident? (0-100%)
3. Combine predictions
4. Remove duplicates

Speed: 30+ images per second (real-time!)

Project: Simple Object Detector

• Simplified YOLO
• Dataset: COCO subset
• Task: Draw boxes around objects

—

👁️ Part 4: Vision Transformers – Attention for Images

New Idea

Remember transformers from Module 4? Use them for images!

How:
1. Split image into patches (16×16 pixels each)
2. Treat each patch like a “word”
3. Use transformer attention
4. Predict image class

Project: Mini Vision Transformer

• Small ViT (10M parameters)
• Dataset: CIFAR-10
• Training: 2-3 hours
• Compare: CNN vs Transformer

—

⚡ ISL Optimization

1. MobileNet Architecture

• Regular convolution: Expensive
• Depthwise separable: 9x faster, same accuracy!

2. Progressive Resizing

Start: 64x64 images (fast!)
Middle: 128x128 images
End: 256x256 images (best quality)

Result: 2x faster training!

3. INT8 Quantization for Inference

• After training, convert to 8-bit
• Result: 4x faster predictions

4. ONNX Runtime

• Export model to ONNX format
• Use optimized runtime
• Result: 2-3x faster inference on CPU

—

📚 Resources

• CIFAR-10, ImageNet subsets
• COCO dataset (object detection)
• Albumentations library
• OpenCV
• Roboflow (create custom datasets)

—

✅ Learning Checklist

• [ ] Build ResNet with skip connections
• [ ] Create data augmentation pipeline
• [ ] Implement object detector
• [ ] Build vision transformer
• [ ] Apply MobileNet optimizations
• [ ] Use progressive resizing
• [ ] Deploy with ONNX Runtime

—

🚀 Next Steps

Module 7: Training Tricks

Learn professional techniques to train faster and better!